├── README.md
├── data_loader.py
├── eigen_test_files.txt
├── eigen_train_files.txt
├── eigen_val_files.txt
├── get_dataset.sh
├── get_depth_maps.sh
├── imgs
    ├── train_image_24.jpg
    └── train_image_28.jpg
├── model.py
├── nyu_dataloader.py
├── nyu_loader.py
├── nyu_train.py
├── prepare_data.py
├── train.py
├── train_dorn.py
├── train_patchgan.py
├── train_unet.py
└── unet.py


/README.md:
--------------------------------------------------------------------------------
 1 | # Refining Monocular Depth Estimation using PatchGAN Priors
 2 | Learning Depth for Autonomous driving from Monocular Images
 3 | 
 4 | ## Network
 5 | The repo presents an approach to refine monocular depth estimates by incorporating a patch discriminator. <br>
 6 | 
 7 | To train, please run
 8 | ```
 9 | python train.py [args]
10 | ```
11 | 
12 | # Results
13 | The final results for KITTI are still pending. For the NYU dataset, it achieves a MSE of 0.03 and L1 Loss of 0.08.
14 | 
15 | Here are a few snapshots
16 | 
17 | ![Sample Image 1](imgs/train_image_28.jpg "NYU sample results")
18 | ![Sample Image 2](imgs/train_image_24.jpg "NYU sample results")
19 | 


--------------------------------------------------------------------------------
/data_loader.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import os
 3 | import glob
 4 | # import cv2
 5 | import torch
 6 | from PIL import Image
 7 | from skimage.util import img_as_float
 8 | from skimage.transform import resize
 9 | 
10 | class DataLoader():
11 |     '''
12 |     raw_data_dir = dir containing extracted raw KITTI data (folder containing 2011-09-26 etc.)
13 |     depth_maps_dir = dir containing extracted depth maps
14 |     '''
15 |     
16 |     def __init__(self, raw_images_path, depth_images_path, mode="train"):
17 |         
18 |         self.mode = mode
19 |         
20 |         if self.mode == "train":
21 |             with open('eigen_train_files.txt', 'r') as f:
22 |                 self.files = f.readlines()
23 |         elif self.mode == "test":
24 |             with open('eigen_test_files.txt', 'r') as f:
25 |                 self.files = f.readlines()
26 |         elif self.mode == "val":
27 |             with open('eigen_val_files.txt', 'r') as f:
28 |                 self.files = f.readlines()
29 |             
30 |         self.data = []
31 |         for l in self.files:
32 |             s = l.split()
33 |             for im in s:
34 |                 self.data.append(raw_images_path + im)
35 | 
36 |         self.imgs, self.labels = [], []
37 |         
38 |         for img_name in self.data:
39 |             img_name = img_name.split('.')[0] + '.png'
40 |             tokens = img_name.split('/')
41 |             path = depth_images_path + 'train/' + tokens[-4] + '/proj_depth/groundtruth/' + tokens[-3] + '/' + tokens[-1]
42 |             path = path.split('.')[0] + '.png'
43 |             if os.path.exists(path) and os.path.exists(img_name):
44 |                 self.imgs.append(img_name)
45 |                 self.labels.append(path)
46 |                 
47 |         print('Found %d Images %d'%(len(self.imgs), len(self.labels)))
48 |     
49 |     def __len__(self):
50 |         return len(self.imgs)
51 |     
52 |     def load_data(self, img_file, label_file):
53 |         x = Image.open(img_file) 
54 |         y = Image.open(label_file)
55 |         x = np.array(x).astype(np.float32)
56 |         y = np.array(y).astype(np.float32)
57 |         x = resize(x, (90, 270), anti_aliasing=True)
58 |         y = resize(y, (90, 270), anti_aliasing=True)
59 |         x = x / 255.0
60 |         y[y<1] = y
61 |         y = np.log(y)
62 |         y = y / np.max(y)
63 |         return x, y
64 |         
65 |     def get_one_batch(self, batch_size = 64):
66 |         images = []
67 |         labels = []
68 | 
69 |         while True:
70 |             idx = np.random.choice(len(self.imgs), batch_size)
71 |             for i in idx:
72 |                 x, y = self.load_data(self.imgs[i], self.labels[i])
73 |                 images.append(x)
74 |                 labels.append(y)
75 |             yield torch.from_numpy(np.array(images)).permute(0,3,1,2), torch.from_numpy(np.array(labels)).permute(0,3,1,2)
76 | 


--------------------------------------------------------------------------------
/eigen_test_files.txt:
--------------------------------------------------------------------------------
  1 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000069.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000069.jpg
  2 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000054.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000054.jpg
  3 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000042.jpg
  4 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000057.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000057.jpg
  5 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000030.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000030.jpg
  6 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000027.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000027.jpg
  7 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000012.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000012.jpg
  8 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000075.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000075.jpg
  9 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000036.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000036.jpg
 10 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000033.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000033.jpg
 11 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000015.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000015.jpg
 12 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000072.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000072.jpg
 13 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000003.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000003.jpg
 14 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000039.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000039.jpg
 15 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000009.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000009.jpg
 16 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000051.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000051.jpg
 17 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000060.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000060.jpg
 18 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000021.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000021.jpg
 19 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000000.jpg
 20 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000024.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000024.jpg
 21 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000045.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000045.jpg
 22 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000018.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000018.jpg
 23 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000048.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000048.jpg
 24 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000006.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000006.jpg
 25 | 2011_09_26/2011_09_26_drive_0002_sync/image_02/data/0000000063.jpg 2011_09_26/2011_09_26_drive_0002_sync/image_03/data/0000000063.jpg
 26 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000000.jpg
 27 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000016.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000016.jpg
 28 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000032.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000032.jpg
 29 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000048.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000048.jpg
 30 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000064.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000064.jpg
 31 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000080.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000080.jpg
 32 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000096.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000096.jpg
 33 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000112.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000112.jpg
 34 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000128.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000128.jpg
 35 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000144.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000144.jpg
 36 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000160.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000160.jpg
 37 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000176.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000176.jpg
 38 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000196.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000196.jpg
 39 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000212.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000212.jpg
 40 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000228.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000228.jpg
 41 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000244.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000244.jpg
 42 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000260.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000260.jpg
 43 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000276.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000276.jpg
 44 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000292.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000292.jpg
 45 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000308.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000308.jpg
 46 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000324.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000324.jpg
 47 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000340.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000340.jpg
 48 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000356.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000356.jpg
 49 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000372.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000372.jpg
 50 | 2011_09_26/2011_09_26_drive_0009_sync/image_02/data/0000000388.jpg 2011_09_26/2011_09_26_drive_0009_sync/image_03/data/0000000388.jpg
 51 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000090.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000090.jpg
 52 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000050.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000050.jpg
 53 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000110.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000110.jpg
 54 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000115.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000115.jpg
 55 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000060.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000060.jpg
 56 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000105.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000105.jpg
 57 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000125.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000125.jpg
 58 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000020.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000020.jpg
 59 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000140.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000140.jpg
 60 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000085.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000085.jpg
 61 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000070.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000070.jpg
 62 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000080.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000080.jpg
 63 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000065.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000065.jpg
 64 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000095.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000095.jpg
 65 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000130.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000130.jpg
 66 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000100.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000100.jpg
 67 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000010.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000010.jpg
 68 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000030.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000030.jpg
 69 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000000.jpg
 70 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000135.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000135.jpg
 71 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000040.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000040.jpg
 72 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000005.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000005.jpg
 73 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000120.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000120.jpg
 74 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000045.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000045.jpg
 75 | 2011_09_26/2011_09_26_drive_0013_sync/image_02/data/0000000035.jpg 2011_09_26/2011_09_26_drive_0013_sync/image_03/data/0000000035.jpg
 76 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000003.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000003.jpg
 77 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000069.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000069.jpg
 78 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000057.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000057.jpg
 79 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000012.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000012.jpg
 80 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000072.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000072.jpg
 81 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000018.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000018.jpg
 82 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000063.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000063.jpg
 83 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000000.jpg
 84 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000084.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000084.jpg
 85 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000015.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000015.jpg
 86 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000066.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000066.jpg
 87 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000006.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000006.jpg
 88 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000048.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000048.jpg
 89 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000060.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000060.jpg
 90 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000009.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000009.jpg
 91 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000033.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000033.jpg
 92 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000021.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000021.jpg
 93 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000075.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000075.jpg
 94 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000027.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000027.jpg
 95 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000045.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000045.jpg
 96 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000078.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000078.jpg
 97 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000036.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000036.jpg
 98 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000051.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000051.jpg
 99 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000054.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000054.jpg
100 | 2011_09_26/2011_09_26_drive_0020_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0020_sync/image_03/data/0000000042.jpg
101 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000018.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000018.jpg
102 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000090.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000090.jpg
103 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000126.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000126.jpg
104 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000378.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000378.jpg
105 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000036.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000036.jpg
106 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000288.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000288.jpg
107 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000198.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000198.jpg
108 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000450.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000450.jpg
109 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000144.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000144.jpg
110 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000072.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000072.jpg
111 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000252.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000252.jpg
112 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000180.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000180.jpg
113 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000432.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000432.jpg
114 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000396.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000396.jpg
115 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000054.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000054.jpg
116 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000468.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000468.jpg
117 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000306.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000306.jpg
118 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000108.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000108.jpg
119 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000162.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000162.jpg
120 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000342.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000342.jpg
121 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000270.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000270.jpg
122 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000414.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000414.jpg
123 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000216.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000216.jpg
124 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000360.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000360.jpg
125 | 2011_09_26/2011_09_26_drive_0023_sync/image_02/data/0000000324.jpg 2011_09_26/2011_09_26_drive_0023_sync/image_03/data/0000000324.jpg
126 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000077.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000077.jpg
127 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000035.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000035.jpg
128 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000091.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000091.jpg
129 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000112.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000112.jpg
130 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000007.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000007.jpg
131 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000175.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000175.jpg
132 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000042.jpg
133 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000098.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000098.jpg
134 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000133.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000133.jpg
135 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000161.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000161.jpg
136 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000014.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000014.jpg
137 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000126.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000126.jpg
138 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000168.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000168.jpg
139 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000070.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000070.jpg
140 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000084.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000084.jpg
141 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000140.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000140.jpg
142 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000049.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000049.jpg
143 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000000.jpg
144 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000182.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000182.jpg
145 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000147.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000147.jpg
146 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000056.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000056.jpg
147 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000063.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000063.jpg
148 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000021.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000021.jpg
149 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000119.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000119.jpg
150 | 2011_09_26/2011_09_26_drive_0027_sync/image_02/data/0000000028.jpg 2011_09_26/2011_09_26_drive_0027_sync/image_03/data/0000000028.jpg
151 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000380.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000380.jpg
152 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000394.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000394.jpg
153 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000324.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000324.jpg
154 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000000.jpg
155 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000268.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000268.jpg
156 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000366.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000366.jpg
157 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000296.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000296.jpg
158 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000014.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000014.jpg
159 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000028.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000028.jpg
160 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000182.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000182.jpg
161 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000168.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000168.jpg
162 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000196.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000196.jpg
163 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000140.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000140.jpg
164 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000084.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000084.jpg
165 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000056.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000056.jpg
166 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000112.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000112.jpg
167 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000352.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000352.jpg
168 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000126.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000126.jpg
169 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000070.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000070.jpg
170 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000310.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000310.jpg
171 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000154.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000154.jpg
172 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000098.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000098.jpg
173 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000408.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000408.jpg
174 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000042.jpg
175 | 2011_09_26/2011_09_26_drive_0029_sync/image_02/data/0000000338.jpg 2011_09_26/2011_09_26_drive_0029_sync/image_03/data/0000000338.jpg
176 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000000.jpg
177 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000128.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000128.jpg
178 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000192.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000192.jpg
179 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000032.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000032.jpg
180 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000352.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000352.jpg
181 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000608.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000608.jpg
182 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000224.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000224.jpg
183 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000576.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000576.jpg
184 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000672.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000672.jpg
185 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000064.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000064.jpg
186 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000448.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000448.jpg
187 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000704.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000704.jpg
188 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000640.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000640.jpg
189 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000512.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000512.jpg
190 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000768.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000768.jpg
191 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000160.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000160.jpg
192 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000416.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000416.jpg
193 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000480.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000480.jpg
194 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000800.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000800.jpg
195 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000288.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000288.jpg
196 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000544.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000544.jpg
197 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000096.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000096.jpg
198 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000384.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000384.jpg
199 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000256.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000256.jpg
200 | 2011_09_26/2011_09_26_drive_0036_sync/image_02/data/0000000320.jpg 2011_09_26/2011_09_26_drive_0036_sync/image_03/data/0000000320.jpg
201 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000000.jpg
202 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000005.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000005.jpg
203 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000010.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000010.jpg
204 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000015.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000015.jpg
205 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000020.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000020.jpg
206 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000025.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000025.jpg
207 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000030.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000030.jpg
208 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000035.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000035.jpg
209 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000040.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000040.jpg
210 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000045.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000045.jpg
211 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000050.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000050.jpg
212 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000055.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000055.jpg
213 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000060.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000060.jpg
214 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000065.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000065.jpg
215 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000070.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000070.jpg
216 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000075.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000075.jpg
217 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000080.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000080.jpg
218 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000085.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000085.jpg
219 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000090.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000090.jpg
220 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000095.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000095.jpg
221 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000100.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000100.jpg
222 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000105.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000105.jpg
223 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000110.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000110.jpg
224 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000115.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000115.jpg
225 | 2011_09_26/2011_09_26_drive_0046_sync/image_02/data/0000000120.jpg 2011_09_26/2011_09_26_drive_0046_sync/image_03/data/0000000120.jpg
226 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000000.jpg
227 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000001.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000001.jpg
228 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000002.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000002.jpg
229 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000003.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000003.jpg
230 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000004.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000004.jpg
231 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000005.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000005.jpg
232 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000006.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000006.jpg
233 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000007.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000007.jpg
234 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000008.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000008.jpg
235 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000009.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000009.jpg
236 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000010.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000010.jpg
237 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000011.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000011.jpg
238 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000012.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000012.jpg
239 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000013.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000013.jpg
240 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000014.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000014.jpg
241 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000015.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000015.jpg
242 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000016.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000016.jpg
243 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000017.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000017.jpg
244 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000018.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000018.jpg
245 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000019.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000019.jpg
246 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000020.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000020.jpg
247 | 2011_09_26/2011_09_26_drive_0048_sync/image_02/data/0000000021.jpg 2011_09_26/2011_09_26_drive_0048_sync/image_03/data/0000000021.jpg
248 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000046.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000046.jpg
249 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000014.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000014.jpg
250 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000036.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000036.jpg
251 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000028.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000028.jpg
252 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000026.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000026.jpg
253 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000050.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000050.jpg
254 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000040.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000040.jpg
255 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000008.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000008.jpg
256 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000016.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000016.jpg
257 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000044.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000044.jpg
258 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000018.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000018.jpg
259 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000032.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000032.jpg
260 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000042.jpg
261 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000010.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000010.jpg
262 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000020.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000020.jpg
263 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000048.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000048.jpg
264 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000052.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000052.jpg
265 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000006.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000006.jpg
266 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000030.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000030.jpg
267 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000012.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000012.jpg
268 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000038.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000038.jpg
269 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000000.jpg
270 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000002.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000002.jpg
271 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000004.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000004.jpg
272 | 2011_09_26/2011_09_26_drive_0052_sync/image_02/data/0000000022.jpg 2011_09_26/2011_09_26_drive_0052_sync/image_03/data/0000000022.jpg
273 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000011.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000011.jpg
274 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000033.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000033.jpg
275 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000242.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000242.jpg
276 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000253.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000253.jpg
277 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000286.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000286.jpg
278 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000154.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000154.jpg
279 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000099.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000099.jpg
280 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000220.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000220.jpg
281 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000022.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000022.jpg
282 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000077.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000077.jpg
283 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000187.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000187.jpg
284 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000143.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000143.jpg
285 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000066.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000066.jpg
286 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000176.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000176.jpg
287 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000110.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000110.jpg
288 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000275.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000275.jpg
289 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000264.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000264.jpg
290 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000198.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000198.jpg
291 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000055.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000055.jpg
292 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000088.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000088.jpg
293 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000121.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000121.jpg
294 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000209.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000209.jpg
295 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000165.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000165.jpg
296 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000231.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000231.jpg
297 | 2011_09_26/2011_09_26_drive_0056_sync/image_02/data/0000000044.jpg 2011_09_26/2011_09_26_drive_0056_sync/image_03/data/0000000044.jpg
298 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000056.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000056.jpg
299 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000000.jpg
300 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000344.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000344.jpg
301 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000358.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000358.jpg
302 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000316.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000316.jpg
303 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000238.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000238.jpg
304 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000098.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000098.jpg
305 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000112.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000112.jpg
306 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000028.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000028.jpg
307 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000014.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000014.jpg
308 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000330.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000330.jpg
309 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000154.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000154.jpg
310 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000042.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000042.jpg
311 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000302.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000302.jpg
312 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000182.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000182.jpg
313 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000288.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000288.jpg
314 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000140.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000140.jpg
315 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000274.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000274.jpg
316 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000224.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000224.jpg
317 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000372.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000372.jpg
318 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000196.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000196.jpg
319 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000126.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000126.jpg
320 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000084.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000084.jpg
321 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000210.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000210.jpg
322 | 2011_09_26/2011_09_26_drive_0059_sync/image_02/data/0000000070.jpg 2011_09_26/2011_09_26_drive_0059_sync/image_03/data/0000000070.jpg
323 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000528.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000528.jpg
324 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000308.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000308.jpg
325 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000044.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000044.jpg
326 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000352.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000352.jpg
327 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000066.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000066.jpg
328 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000000.jpg
329 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000506.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000506.jpg
330 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000176.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000176.jpg
331 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000022.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000022.jpg
332 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000242.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000242.jpg
333 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000462.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000462.jpg
334 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000418.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000418.jpg
335 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000110.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000110.jpg
336 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000440.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000440.jpg
337 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000396.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000396.jpg
338 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000154.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000154.jpg
339 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000374.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000374.jpg
340 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000088.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000088.jpg
341 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000286.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000286.jpg
342 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000550.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000550.jpg
343 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000264.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000264.jpg
344 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000220.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000220.jpg
345 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000330.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000330.jpg
346 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000484.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000484.jpg
347 | 2011_09_26/2011_09_26_drive_0064_sync/image_02/data/0000000198.jpg 2011_09_26/2011_09_26_drive_0064_sync/image_03/data/0000000198.jpg
348 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000283.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000283.jpg
349 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000361.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000361.jpg
350 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000270.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000270.jpg
351 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000127.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000127.jpg
352 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000205.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000205.jpg
353 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000218.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000218.jpg
354 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000153.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000153.jpg
355 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000335.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000335.jpg
356 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000192.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000192.jpg
357 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000348.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000348.jpg
358 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000101.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000101.jpg
359 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000049.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000049.jpg
360 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000179.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000179.jpg
361 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000140.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000140.jpg
362 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000374.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000374.jpg
363 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000322.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000322.jpg
364 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000309.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000309.jpg
365 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000244.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000244.jpg
366 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000062.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000062.jpg
367 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000257.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000257.jpg
368 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000088.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000088.jpg
369 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000114.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000114.jpg
370 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000075.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000075.jpg
371 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000296.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000296.jpg
372 | 2011_09_26/2011_09_26_drive_0084_sync/image_02/data/0000000231.jpg 2011_09_26/2011_09_26_drive_0084_sync/image_03/data/0000000231.jpg
373 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000007.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000007.jpg
374 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000196.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000196.jpg
375 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000439.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000439.jpg
376 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000169.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000169.jpg
377 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000115.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000115.jpg
378 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000034.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000034.jpg
379 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000304.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000304.jpg
380 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000331.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000331.jpg
381 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000277.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000277.jpg
382 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000520.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000520.jpg
383 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000682.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000682.jpg
384 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000628.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000628.jpg
385 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000088.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000088.jpg
386 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000601.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000601.jpg
387 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000574.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000574.jpg
388 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000223.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000223.jpg
389 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000655.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000655.jpg
390 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000358.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000358.jpg
391 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000412.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000412.jpg
392 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000142.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000142.jpg
393 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000385.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000385.jpg
394 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000061.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000061.jpg
395 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000493.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000493.jpg
396 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000466.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000466.jpg
397 | 2011_09_26/2011_09_26_drive_0086_sync/image_02/data/0000000250.jpg 2011_09_26/2011_09_26_drive_0086_sync/image_03/data/0000000250.jpg
398 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000000.jpg
399 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000016.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000016.jpg
400 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000032.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000032.jpg
401 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000048.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000048.jpg
402 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000064.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000064.jpg
403 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000080.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000080.jpg
404 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000096.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000096.jpg
405 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000112.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000112.jpg
406 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000128.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000128.jpg
407 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000144.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000144.jpg
408 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000160.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000160.jpg
409 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000176.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000176.jpg
410 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000192.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000192.jpg
411 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000208.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000208.jpg
412 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000224.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000224.jpg
413 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000240.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000240.jpg
414 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000256.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000256.jpg
415 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000305.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000305.jpg
416 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000321.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000321.jpg
417 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000337.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000337.jpg
418 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000353.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000353.jpg
419 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000369.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000369.jpg
420 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000385.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000385.jpg
421 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000401.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000401.jpg
422 | 2011_09_26/2011_09_26_drive_0093_sync/image_02/data/0000000417.jpg 2011_09_26/2011_09_26_drive_0093_sync/image_03/data/0000000417.jpg
423 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000000.jpg
424 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000019.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000019.jpg
425 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000038.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000038.jpg
426 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000057.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000057.jpg
427 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000076.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000076.jpg
428 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000095.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000095.jpg
429 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000114.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000114.jpg
430 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000133.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000133.jpg
431 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000152.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000152.jpg
432 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000171.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000171.jpg
433 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000190.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000190.jpg
434 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000209.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000209.jpg
435 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000228.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000228.jpg
436 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000247.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000247.jpg
437 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000266.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000266.jpg
438 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000285.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000285.jpg
439 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000304.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000304.jpg
440 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000323.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000323.jpg
441 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000342.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000342.jpg
442 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000361.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000361.jpg
443 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000380.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000380.jpg
444 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000399.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000399.jpg
445 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000418.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000418.jpg
446 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000437.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000437.jpg
447 | 2011_09_26/2011_09_26_drive_0096_sync/image_02/data/0000000456.jpg 2011_09_26/2011_09_26_drive_0096_sync/image_03/data/0000000456.jpg
448 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000692.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000692.jpg
449 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000930.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000930.jpg
450 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000760.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000760.jpg
451 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000896.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000896.jpg
452 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000284.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000284.jpg
453 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000148.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000148.jpg
454 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000522.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000522.jpg
455 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000794.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000794.jpg
456 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000624.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000624.jpg
457 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000726.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000726.jpg
458 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000216.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000216.jpg
459 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000318.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000318.jpg
460 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000488.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000488.jpg
461 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000590.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000590.jpg
462 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000454.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000454.jpg
463 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000862.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000862.jpg
464 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000386.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000386.jpg
465 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000352.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000352.jpg
466 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000420.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000420.jpg
467 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000658.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000658.jpg
468 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000828.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000828.jpg
469 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000556.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000556.jpg
470 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000114.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000114.jpg
471 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000182.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000182.jpg
472 | 2011_09_26/2011_09_26_drive_0101_sync/image_02/data/0000000080.jpg 2011_09_26/2011_09_26_drive_0101_sync/image_03/data/0000000080.jpg
473 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000015.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000015.jpg
474 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000035.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000035.jpg
475 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000043.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000043.jpg
476 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000051.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000051.jpg
477 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000059.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000059.jpg
478 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000067.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000067.jpg
479 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000075.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000075.jpg
480 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000083.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000083.jpg
481 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000091.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000091.jpg
482 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000099.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000099.jpg
483 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000107.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000107.jpg
484 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000115.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000115.jpg
485 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000123.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000123.jpg
486 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000131.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000131.jpg
487 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000139.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000139.jpg
488 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000147.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000147.jpg
489 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000155.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000155.jpg
490 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000163.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000163.jpg
491 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000171.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000171.jpg
492 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000179.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000179.jpg
493 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000187.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000187.jpg
494 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000195.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000195.jpg
495 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000203.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000203.jpg
496 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000211.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000211.jpg
497 | 2011_09_26/2011_09_26_drive_0106_sync/image_02/data/0000000219.jpg 2011_09_26/2011_09_26_drive_0106_sync/image_03/data/0000000219.jpg
498 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000312.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000312.jpg
499 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000494.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000494.jpg
500 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000104.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000104.jpg
501 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000130.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000130.jpg
502 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000156.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000156.jpg
503 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000182.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000182.jpg
504 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000598.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000598.jpg
505 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000416.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000416.jpg
506 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000364.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000364.jpg
507 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000026.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000026.jpg
508 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000078.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000078.jpg
509 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000572.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000572.jpg
510 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000468.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000468.jpg
511 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000260.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000260.jpg
512 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000624.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000624.jpg
513 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000234.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000234.jpg
514 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000442.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000442.jpg
515 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000390.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000390.jpg
516 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000546.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000546.jpg
517 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000286.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000286.jpg
518 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000000.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000000.jpg
519 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000338.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000338.jpg
520 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000208.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000208.jpg
521 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000650.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000650.jpg
522 | 2011_09_26/2011_09_26_drive_0117_sync/image_02/data/0000000052.jpg 2011_09_26/2011_09_26_drive_0117_sync/image_03/data/0000000052.jpg
523 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000024.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000024.jpg
524 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000021.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000021.jpg
525 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000036.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000036.jpg
526 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000000.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000000.jpg
527 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000051.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000051.jpg
528 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000018.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000018.jpg
529 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000033.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000033.jpg
530 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000090.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000090.jpg
531 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000045.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000045.jpg
532 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000054.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000054.jpg
533 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000012.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000012.jpg
534 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000039.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000039.jpg
535 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000009.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000009.jpg
536 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000003.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000003.jpg
537 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000030.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000030.jpg
538 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000078.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000078.jpg
539 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000060.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000060.jpg
540 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000048.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000048.jpg
541 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000084.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000084.jpg
542 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000081.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000081.jpg
543 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000006.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000006.jpg
544 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000057.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000057.jpg
545 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000072.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000072.jpg
546 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000087.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000087.jpg
547 | 2011_09_28/2011_09_28_drive_0002_sync/image_02/data/0000000063.jpg 2011_09_28/2011_09_28_drive_0002_sync/image_03/data/0000000063.jpg
548 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000252.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000252.jpg
549 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000540.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000540.jpg
550 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000001054.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000001054.jpg
551 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000036.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000036.jpg
552 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000360.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000360.jpg
553 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000807.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000807.jpg
554 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000879.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000879.jpg
555 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000288.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000288.jpg
556 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000771.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000771.jpg
557 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000000.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000000.jpg
558 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000216.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000216.jpg
559 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000951.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000951.jpg
560 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000324.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000324.jpg
561 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000432.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000432.jpg
562 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000504.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000504.jpg
563 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000576.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000576.jpg
564 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000108.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000108.jpg
565 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000180.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000180.jpg
566 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000072.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000072.jpg
567 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000612.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000612.jpg
568 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000915.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000915.jpg
569 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000735.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000735.jpg
570 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000144.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000144.jpg
571 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000396.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000396.jpg
572 | 2011_09_29/2011_09_29_drive_0071_sync/image_02/data/0000000468.jpg 2011_09_29/2011_09_29_drive_0071_sync/image_03/data/0000000468.jpg
573 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000132.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000132.jpg
574 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000011.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000011.jpg
575 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000154.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000154.jpg
576 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000022.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000022.jpg
577 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000242.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000242.jpg
578 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000198.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000198.jpg
579 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000176.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000176.jpg
580 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000231.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000231.jpg
581 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000275.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000275.jpg
582 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000220.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000220.jpg
583 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000088.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000088.jpg
584 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000143.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000143.jpg
585 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000055.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000055.jpg
586 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000033.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000033.jpg
587 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000187.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000187.jpg
588 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000110.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000110.jpg
589 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000044.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000044.jpg
590 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000077.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000077.jpg
591 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000066.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000066.jpg
592 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000000.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000000.jpg
593 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000165.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000165.jpg
594 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000264.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000264.jpg
595 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000253.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000253.jpg
596 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000209.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000209.jpg
597 | 2011_09_30/2011_09_30_drive_0016_sync/image_02/data/0000000121.jpg 2011_09_30/2011_09_30_drive_0016_sync/image_03/data/0000000121.jpg
598 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000107.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000107.jpg
599 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002247.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002247.jpg
600 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001391.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001391.jpg
601 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000535.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000535.jpg
602 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001819.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001819.jpg
603 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001177.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001177.jpg
604 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000428.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000428.jpg
605 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001926.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001926.jpg
606 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000749.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000749.jpg
607 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001284.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001284.jpg
608 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002140.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002140.jpg
609 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001605.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001605.jpg
610 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001498.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001498.jpg
611 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000642.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000642.jpg
612 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002740.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002740.jpg
613 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002419.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002419.jpg
614 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000856.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000856.jpg
615 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002526.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002526.jpg
616 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001712.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001712.jpg
617 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000001070.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000001070.jpg
618 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000000.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000000.jpg
619 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002033.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002033.jpg
620 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000214.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000214.jpg
621 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000000963.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000000963.jpg
622 | 2011_09_30/2011_09_30_drive_0018_sync/image_02/data/0000002633.jpg 2011_09_30/2011_09_30_drive_0018_sync/image_03/data/0000002633.jpg
623 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000533.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000533.jpg
624 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000001040.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000001040.jpg
625 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000082.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000082.jpg
626 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000205.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000205.jpg
627 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000835.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000835.jpg
628 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000451.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000451.jpg
629 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000164.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000164.jpg
630 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000794.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000794.jpg
631 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000328.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000328.jpg
632 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000615.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000615.jpg
633 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000917.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000917.jpg
634 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000369.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000369.jpg
635 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000287.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000287.jpg
636 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000123.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000123.jpg
637 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000876.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000876.jpg
638 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000410.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000410.jpg
639 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000492.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000492.jpg
640 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000958.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000958.jpg
641 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000656.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000656.jpg
642 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000000.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000000.jpg
643 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000753.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000753.jpg
644 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000574.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000574.jpg
645 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000001081.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000001081.jpg
646 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000041.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000041.jpg
647 | 2011_09_30/2011_09_30_drive_0027_sync/image_02/data/0000000246.jpg 2011_09_30/2011_09_30_drive_0027_sync/image_03/data/0000000246.jpg
648 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000002906.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000002906.jpg
649 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000002544.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000002544.jpg
650 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000362.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000362.jpg
651 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000004535.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000004535.jpg
652 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000734.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000734.jpg
653 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000001096.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000001096.jpg
654 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000004173.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000004173.jpg
655 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000543.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000543.jpg
656 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000001277.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000001277.jpg
657 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000004354.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000004354.jpg
658 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000001458.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000001458.jpg
659 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000001820.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000001820.jpg
660 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003449.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003449.jpg
661 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003268.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003268.jpg
662 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000915.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000915.jpg
663 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000002363.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000002363.jpg
664 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000002725.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000002725.jpg
665 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000181.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000181.jpg
666 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000001639.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000001639.jpg
667 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003992.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003992.jpg
668 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003087.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003087.jpg
669 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000002001.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000002001.jpg
670 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003811.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003811.jpg
671 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000003630.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000003630.jpg
672 | 2011_10_03/2011_10_03_drive_0027_sync/image_02/data/0000000000.jpg 2011_10_03/2011_10_03_drive_0027_sync/image_03/data/0000000000.jpg
673 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000096.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000096.jpg
674 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000800.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000800.jpg
675 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000320.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000320.jpg
676 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000576.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000576.jpg
677 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000000.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000000.jpg
678 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000480.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000480.jpg
679 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000640.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000640.jpg
680 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000032.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000032.jpg
681 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000384.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000384.jpg
682 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000160.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000160.jpg
683 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000704.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000704.jpg
684 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000736.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000736.jpg
685 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000672.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000672.jpg
686 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000064.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000064.jpg
687 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000288.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000288.jpg
688 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000352.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000352.jpg
689 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000512.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000512.jpg
690 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000544.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000544.jpg
691 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000608.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000608.jpg
692 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000128.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000128.jpg
693 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000224.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000224.jpg
694 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000416.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000416.jpg
695 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000192.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000192.jpg
696 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000448.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000448.jpg
697 | 2011_10_03/2011_10_03_drive_0047_sync/image_02/data/0000000768.jpg 2011_10_03/2011_10_03_drive_0047_sync/image_03/data/0000000768.jpg
698 | 


--------------------------------------------------------------------------------
/get_dataset.sh:
--------------------------------------------------------------------------------
  1 | #!/bin/bash
  2 | 
  3 | files=(
  4 | 2011_09_26_calib.zip
  5 | 2011_09_26_drive_0001
  6 | 2011_09_26_drive_0002
  7 | 2011_09_26_drive_0005
  8 | 2011_09_26_drive_0009
  9 | 2011_09_26_drive_0011
 10 | 2011_09_26_drive_0013
 11 | 2011_09_26_drive_0014
 12 | 2011_09_26_drive_0015
 13 | 2011_09_26_drive_0017
 14 | 2011_09_26_drive_0018
 15 | 2011_09_26_drive_0019
 16 | 2011_09_26_drive_0020
 17 | 2011_09_26_drive_0022
 18 | 2011_09_26_drive_0023
 19 | 2011_09_26_drive_0027
 20 | 2011_09_26_drive_0028
 21 | 2011_09_26_drive_0029
 22 | 2011_09_26_drive_0032
 23 | 2011_09_26_drive_0035
 24 | 2011_09_26_drive_0036
 25 | 2011_09_26_drive_0039
 26 | 2011_09_26_drive_0046
 27 | 2011_09_26_drive_0048
 28 | 2011_09_26_drive_0051
 29 | 2011_09_26_drive_0052
 30 | 2011_09_26_drive_0056
 31 | 2011_09_26_drive_0057
 32 | 2011_09_26_drive_0059
 33 | 2011_09_26_drive_0060
 34 | 2011_09_26_drive_0061
 35 | 2011_09_26_drive_0064
 36 | 2011_09_26_drive_0070
 37 | 2011_09_26_drive_0079
 38 | 2011_09_26_drive_0084
 39 | 2011_09_26_drive_0086
 40 | 2011_09_26_drive_0087
 41 | 2011_09_26_drive_0091
 42 | 2011_09_26_drive_0093
 43 | 2011_09_26_drive_0095
 44 | 2011_09_26_drive_0096
 45 | 2011_09_26_drive_0101
 46 | 2011_09_26_drive_0104
 47 | 2011_09_26_drive_0106
 48 | 2011_09_26_drive_0113
 49 | 2011_09_26_drive_0117
 50 | 2011_09_26_drive_0119
 51 | 2011_09_28_calib.zip
 52 | 2011_09_28_drive_0001
 53 | 2011_09_28_drive_0002
 54 | 2011_09_28_drive_0016
 55 | 2011_09_28_drive_0021
 56 | 2011_09_28_drive_0034
 57 | 2011_09_28_drive_0035
 58 | 2011_09_28_drive_0037
 59 | 2011_09_28_drive_0038
 60 | 2011_09_28_drive_0039
 61 | 2011_09_28_drive_0043
 62 | 2011_09_28_drive_0045
 63 | 2011_09_28_drive_0047
 64 | 2011_09_28_drive_0053
 65 | 2011_09_28_drive_0054
 66 | 2011_09_28_drive_0057
 67 | 2011_09_28_drive_0065
 68 | 2011_09_28_drive_0066
 69 | 2011_09_28_drive_0068
 70 | 2011_09_28_drive_0070
 71 | 2011_09_28_drive_0071
 72 | 2011_09_28_drive_0075
 73 | 2011_09_28_drive_0077
 74 | 2011_09_28_drive_0078
 75 | 2011_09_28_drive_0080
 76 | 2011_09_28_drive_0082
 77 | 2011_09_28_drive_0086
 78 | 2011_09_28_drive_0087
 79 | 2011_09_28_drive_0089
 80 | 2011_09_28_drive_0090
 81 | 2011_09_28_drive_0094
 82 | 2011_09_28_drive_0095
 83 | 2011_09_28_drive_0096
 84 | 2011_09_28_drive_0098
 85 | 2011_09_28_drive_0100
 86 | 2011_09_28_drive_0102
 87 | 2011_09_28_drive_0103
 88 | 2011_09_28_drive_0104
 89 | 2011_09_28_drive_0106
 90 | 2011_09_28_drive_0108
 91 | 2011_09_28_drive_0110
 92 | 2011_09_28_drive_0113
 93 | 2011_09_28_drive_0117
 94 | 2011_09_28_drive_0119
 95 | 2011_09_28_drive_0121
 96 | 2011_09_28_drive_0122
 97 | 2011_09_28_drive_0125
 98 | 2011_09_28_drive_0126
 99 | 2011_09_28_drive_0128
100 | 2011_09_28_drive_0132
101 | 2011_09_28_drive_0134
102 | 2011_09_28_drive_0135
103 | 2011_09_28_drive_0136
104 | 2011_09_28_drive_0138
105 | 2011_09_28_drive_0141
106 | 2011_09_28_drive_0143
107 | 2011_09_28_drive_0145
108 | 2011_09_28_drive_0146
109 | 2011_09_28_drive_0149
110 | 2011_09_28_drive_0153
111 | 2011_09_28_drive_0154
112 | 2011_09_28_drive_0155
113 | 2011_09_28_drive_0156
114 | 2011_09_28_drive_0160
115 | 2011_09_28_drive_0161
116 | 2011_09_28_drive_0162
117 | 2011_09_28_drive_0165
118 | 2011_09_28_drive_0166
119 | 2011_09_28_drive_0167
120 | 2011_09_28_drive_0168
121 | 2011_09_28_drive_0171
122 | 2011_09_28_drive_0174
123 | 2011_09_28_drive_0177
124 | 2011_09_28_drive_0179
125 | 2011_09_28_drive_0183
126 | 2011_09_28_drive_0184
127 | 2011_09_28_drive_0185
128 | 2011_09_28_drive_0186
129 | 2011_09_28_drive_0187
130 | 2011_09_28_drive_0191
131 | 2011_09_28_drive_0192
132 | 2011_09_28_drive_0195
133 | 2011_09_28_drive_0198
134 | 2011_09_28_drive_0199
135 | 2011_09_28_drive_0201
136 | 2011_09_28_drive_0204
137 | 2011_09_28_drive_0205
138 | 2011_09_28_drive_0208
139 | 2011_09_28_drive_0209
140 | 2011_09_28_drive_0214
141 | 2011_09_28_drive_0216
142 | 2011_09_28_drive_0220
143 | 2011_09_28_drive_0222
144 | 2011_09_28_drive_0225
145 | 2011_09_29_calib.zip
146 | 2011_09_29_drive_0004
147 | 2011_09_29_drive_0026
148 | 2011_09_29_drive_0071
149 | 2011_09_29_drive_0108
150 | 2011_09_30_calib.zip
151 | 2011_09_30_drive_0016
152 | 2011_09_30_drive_0018
153 | 2011_09_30_drive_0020
154 | 2011_09_30_drive_0027
155 | 2011_09_30_drive_0028
156 | 2011_09_30_drive_0033
157 | 2011_09_30_drive_0034
158 | 2011_09_30_drive_0072
159 | 2011_10_03_calib.zip
160 | 2011_10_03_drive_0027
161 | 2011_10_03_drive_0034
162 | 2011_10_03_drive_0042
163 | 2011_10_03_drive_0047
164 | 2011_10_03_drive_0058
165 | )
166 | 
167 | for i in ${files[@]}; do
168 |         if [ ${i:(-3)} != "zip" ]
169 |         then
170 |                 shortname=$i'_sync.zip'
171 |                 fullname=$i'/'$i'_sync.zip'
172 |         else
173 |                 shortname=$i
174 |                 fullname=$i
175 |         fi
176 | 	echo "Downloading: "$shortname
177 |         wget -c 'https://s3.eu-central-1.amazonaws.com/avg-kitti/raw_data/'$fullname
178 |         unzip -o $shortname
179 |         rm $shortname
180 | done
181 | 


--------------------------------------------------------------------------------
/get_depth_maps.sh:
--------------------------------------------------------------------------------
1 | wget -c https://s3.eu-central-1.amazonaws.com/avg-kitti/data_depth_annotated.zip
2 | 


--------------------------------------------------------------------------------
/imgs/train_image_24.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/manjotms10/PatchGAN/83de60576137f958efcd1eaa66e8cb62001887aa/imgs/train_image_24.jpg


--------------------------------------------------------------------------------
/imgs/train_image_28.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/manjotms10/PatchGAN/83de60576137f958efcd1eaa66e8cb62001887aa/imgs/train_image_28.jpg


--------------------------------------------------------------------------------
/model.py:
--------------------------------------------------------------------------------
  1 | import torch
  2 | import torch.nn as nn
  3 | import torch.utils.model_zoo as model_zoo
  4 | import CSPN as post_process
  5 | import update_model
  6 | 
  7 | # memory analyze
  8 | import gc
  9 | 
 10 | __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
 11 |            'resnet152']
 12 | 
 13 | model_urls = {
 14 |     'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
 15 |     'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
 16 |     'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
 17 |     'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
 18 |     'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
 19 | }
 20 | 
 21 | model_path = {
 22 |     'resnet18': 'pretrained/resnet18.pth',
 23 |     'resnet50': 'pretrained/resnet50.pth'
 24 | }
 25 | 
 26 | 
 27 | def conv3x3(in_planes, out_planes, stride=1):
 28 |     """3x3 convolution with padding"""
 29 |     return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
 30 |                      padding=1, bias=False)
 31 | 
 32 | 
 33 | class BasicBlock(nn.Module):
 34 |     expansion = 1
 35 | 
 36 |     def __init__(self, inplanes, planes, stride=1, downsample=None):
 37 |         super(BasicBlock, self).__init__()
 38 |         self.conv1 = conv3x3(inplanes, planes, stride)
 39 |         self.bn1 = nn.BatchNorm2d(planes)
 40 |         self.relu = nn.ReLU(inplace=True)
 41 |         self.conv2 = conv3x3(planes, planes)
 42 |         self.bn2 = nn.BatchNorm2d(planes)
 43 |         self.downsample = downsample
 44 |         self.stride = stride
 45 | 
 46 |     def forward(self, x):
 47 |         residual = x
 48 | 
 49 |         out = self.conv1(x)
 50 |         out = self.bn1(out)
 51 |         out = self.relu(out)
 52 | 
 53 |         out = self.conv2(out)
 54 |         out = self.bn2(out)
 55 | 
 56 |         if self.downsample is not None:
 57 |             residual = self.downsample(x)
 58 | 
 59 |         out += residual
 60 |         out = self.relu(out)
 61 | 
 62 |         return out
 63 | 
 64 | 
 65 | class Bottleneck(nn.Module):
 66 |     expansion = 4
 67 | 
 68 |     def __init__(self, inplanes, planes, stride=1, downsample=None):
 69 |         super(Bottleneck, self).__init__()
 70 |         self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
 71 |         self.bn1 = nn.BatchNorm2d(planes)
 72 |         self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
 73 |                                padding=1, bias=False)
 74 |         self.bn2 = nn.BatchNorm2d(planes)
 75 |         self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
 76 |         self.bn3 = nn.BatchNorm2d(planes * 4)
 77 |         self.relu = nn.ReLU(inplace=True)
 78 |         self.downsample = downsample
 79 |         self.stride = stride
 80 | 
 81 |     def forward(self, x):
 82 |         residual = x
 83 | 
 84 |         out = self.conv1(x)
 85 |         out = self.bn1(out)
 86 |         out = self.relu(out)
 87 | 
 88 |         out = self.conv2(out)
 89 |         out = self.bn2(out)
 90 |         out = self.relu(out)
 91 | 
 92 |         out = self.conv3(out)
 93 |         out = self.bn3(out)
 94 | 
 95 |         if self.downsample is not None:
 96 |             residual = self.downsample(x)
 97 | 
 98 |         out += residual
 99 |         out = self.relu(out)
100 | 
101 |         return out
102 | 
103 | 
104 | class UpProj_Block(nn.Module):
105 |     def __init__(self, in_channels, out_channels, oheight=0, owidth=0):
106 |         super(UpProj_Block, self).__init__()
107 |         self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
108 |         self.bn1 = nn.BatchNorm2d(out_channels)
109 |         self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
110 |         self.bn2 = nn.BatchNorm2d(out_channels)
111 |         self.sc_conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
112 |         self.sc_bn1 = nn.BatchNorm2d(out_channels)
113 |         self.relu = nn.ReLU(inplace=True)
114 |         self.oheight = oheight
115 |         self.owidth = owidth
116 | 
117 |     def _up_pooling(self, x, scale):
118 |         oheight = 0
119 |         owidth = 0
120 |         if self.oheight == 0 and self.owidth == 0:
121 |             oheight = scale * x.size(2)
122 |             owidth = scale * x.size(3)
123 |             x = nn.Upsample(scale_factor=scale, mode='nearest')(x)
124 |         else:
125 |             oheight = self.oheight
126 |             owidth = self.owidth
127 |             x = nn.Upsample(size=(oheight, owidth), mode='nearest')(x)
128 |         mask = torch.zeros_like(x)
129 |         for h in range(0, oheight, 2):
130 |             for w in range(0, owidth, 2):
131 |                 mask[:, :, h, w] = 1
132 |         x = torch.mul(mask, x)
133 |         return x
134 | 
135 |     def forward(self, x):
136 |         x = self._up_pooling(x, 2)
137 |         out = self.relu(self.bn1(self.conv1(x)))
138 |         out = self.bn2(self.conv2(out))
139 |         short_cut = self.sc_bn1(self.sc_conv1(x))
140 |         out += short_cut
141 |         out = self.relu(out)
142 |         return out
143 | 
144 | 
145 | class Simple_Gudi_UpConv_Block(nn.Module):
146 |     def __init__(self, in_channels, out_channels, oheight=0, owidth=0):
147 |         super(Simple_Gudi_UpConv_Block, self).__init__()
148 |         self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
149 |         self.bn1 = nn.BatchNorm2d(out_channels)
150 |         self.relu = nn.ReLU(inplace=True)
151 |         self.oheight = oheight
152 |         self.owidth = owidth
153 | 
154 |     def _up_pooling(self, x, scale):
155 | 
156 |         x = nn.Upsample(scale_factor=scale, mode='nearest')(x)
157 |         if self.oheight != 0 and self.owidth != 0:
158 |             x = x.narrow(2, 0, self.oheight)
159 |             x = x.narrow(3, 0, self.owidth)
160 |         mask = torch.zeros_like(x)
161 |         for h in range(0, self.oheight, 2):
162 |             for w in range(0, self.owidth, 2):
163 |                 mask[:, :, h, w] = 1
164 |         x = torch.mul(mask, x)
165 |         return x
166 | 
167 |     def forward(self, x):
168 |         x = self._up_pooling(x, 2)
169 |         out = self.relu(self.bn1(self.conv1(x)))
170 |         return out
171 | 
172 | 
173 | class Simple_Gudi_UpConv_Block_Last_Layer(nn.Module):
174 |     def __init__(self, in_channels, out_channels, oheight=0, owidth=0):
175 |         super(Simple_Gudi_UpConv_Block_Last_Layer, self).__init__()
176 |         self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
177 |         self.oheight = oheight
178 |         self.owidth = owidth
179 | 
180 |     def _up_pooling(self, x, scale):
181 | 
182 |         x = nn.Upsample(scale_factor=scale, mode='nearest')(x)
183 |         if self.oheight != 0 and self.owidth != 0:
184 |             x = x.narrow(2, 0, self.oheight)
185 |             x = x.narrow(3, 0, self.owidth)
186 |         mask = torch.zeros_like(x)
187 |         for h in range(0, self.oheight, 2):
188 |             for w in range(0, self.owidth, 2):
189 |                 mask[:, :, h, w] = 1
190 |         x = torch.mul(mask, x)
191 |         return x
192 | 
193 |     def forward(self, x):
194 |         x = self._up_pooling(x, 2)
195 |         out = self.conv1(x)
196 |         return out
197 | 
198 | 
199 | class Gudi_UpProj_Block(nn.Module):
200 |     def __init__(self, in_channels, out_channels, oheight=0, owidth=0):
201 |         super(Gudi_UpProj_Block, self).__init__()
202 |         self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
203 |         self.bn1 = nn.BatchNorm2d(out_channels)
204 |         self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
205 |         self.bn2 = nn.BatchNorm2d(out_channels)
206 |         self.sc_conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
207 |         self.sc_bn1 = nn.BatchNorm2d(out_channels)
208 |         self.relu = nn.ReLU(inplace=True)
209 |         self.oheight = oheight
210 |         self.owidth = owidth
211 | 
212 |     def _up_pooling(self, x, scale):
213 | 
214 |         x = nn.Upsample(scale_factor=scale, mode='nearest')(x)
215 |         if self.oheight != 0 and self.owidth != 0:
216 |             x = x[:, :, 0:self.oheight, 0:self.owidth]
217 |         mask = torch.zeros_like(x)
218 |         for h in range(0, self.oheight, 2):
219 |             for w in range(0, self.owidth, 2):
220 |                 mask[:, :, h, w] = 1
221 |         x = torch.mul(mask, x)
222 |         return x
223 | 
224 |     def forward(self, x):
225 |         x = self._up_pooling(x, 2)
226 |         out = self.relu(self.bn1(self.conv1(x)))
227 |         out = self.bn2(self.conv2(out))
228 |         short_cut = self.sc_bn1(self.sc_conv1(x))
229 |         out += short_cut
230 |         out = self.relu(out)
231 |         return out
232 | 
233 | 
234 | class Gudi_UpProj_Block_Cat(nn.Module):
235 |     def __init__(self, in_channels, out_channels, oheight=0, owidth=0):
236 |         super(Gudi_UpProj_Block_Cat, self).__init__()
237 |         self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
238 |         self.bn1 = nn.BatchNorm2d(out_channels)
239 |         self.conv1_1 = nn.Conv2d(out_channels * 2, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
240 |         self.bn1_1 = nn.BatchNorm2d(out_channels)
241 |         self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
242 |         self.bn2 = nn.BatchNorm2d(out_channels)
243 |         self.sc_conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)
244 |         self.sc_bn1 = nn.BatchNorm2d(out_channels)
245 |         self.relu = nn.ReLU(inplace=True)
246 |         self.oheight = oheight
247 |         self.owidth = owidth
248 | 
249 |     def _up_pooling(self, x, scale):
250 | 
251 |         x = nn.Upsample(scale_factor=scale, mode='nearest')(x)
252 |         if self.oheight != 0 and self.owidth != 0:
253 |             x = x[:, :, 0:self.oheight, 0:self.owidth]
254 |         mask = torch.zeros_like(x)
255 |         for h in range(0, self.oheight, 2):
256 |             for w in range(0, self.owidth, 2):
257 |                 mask[:, :, h, w] = 1
258 |         x = torch.mul(mask, x)
259 |         return x
260 | 
261 |     def forward(self, x, side_input):
262 |         x = self._up_pooling(x, 2)
263 |         out = self.relu(self.bn1(self.conv1(x)))
264 |         out = torch.cat((out, side_input), 1)
265 |         out = self.relu(self.bn1_1(self.conv1_1(out)))
266 |         out = self.bn2(self.conv2(out))
267 |         short_cut = self.sc_bn1(self.sc_conv1(x))
268 |         out += short_cut
269 |         out = self.relu(out)
270 |         return out
271 | 
272 | 
273 | class ResNet(nn.Module):
274 | 
275 |     def __init__(self, block, layers, up_proj_block):
276 |         self.inplanes = 64
277 |         super(ResNet, self).__init__()
278 |         self.conv1_1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
279 |                                  bias=False)
280 |         self.bn1 = nn.BatchNorm2d(64)
281 |         self.relu = nn.ReLU(inplace=True)
282 |         self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
283 |         self.layer1 = self._make_layer(block, 64, layers[0])
284 |         self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
285 |         self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
286 |         self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
287 | 
288 |         self.mid_channel = 256 * block.expansion
289 |         self.conv2 = nn.Conv2d(512 * block.expansion, 512 * block.expansion, kernel_size=3, stride=1, padding=1,
290 |                                bias=False)
291 |         self.bn2 = nn.BatchNorm2d(512 * block.expansion)
292 |         self.conv3 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1, bias=False)
293 | 
294 |         self.post_process_layer = self._make_post_process_layer()
295 | 
296 |         self.gud_up_proj_layer1 = self._make_gud_up_conv_layer(Gudi_UpProj_Block, 2048, 1024, 15, 19)
297 |         self.gud_up_proj_layer2 = self._make_gud_up_conv_layer(Gudi_UpProj_Block_Cat, 1024, 512, 29, 38)
298 |         self.gud_up_proj_layer3 = self._make_gud_up_conv_layer(Gudi_UpProj_Block_Cat, 512, 256, 57, 76)
299 |         self.gud_up_proj_layer4 = self._make_gud_up_conv_layer(Gudi_UpProj_Block_Cat, 256, 64, 114, 152)
300 |         self.gud_up_proj_layer5 = self._make_gud_up_conv_layer(Simple_Gudi_UpConv_Block_Last_Layer, 64, 1, 228, 304)
301 |         self.gud_up_proj_layer6 = self._make_gud_up_conv_layer(Simple_Gudi_UpConv_Block_Last_Layer, 64, 1, 228, 304)
302 | 
303 |     def _make_layer(self, block, planes, blocks, stride=1):
304 |         downsample = None
305 |         if stride != 1 or self.inplanes != planes * block.expansion:
306 |             downsample = nn.Sequential(
307 |                 nn.Conv2d(self.inplanes, planes * block.expansion,
308 |                           kernel_size=1, stride=stride, bias=False),
309 |                 nn.BatchNorm2d(planes * block.expansion),
310 |             )
311 | 
312 |         layers = []
313 |         layers.append(block(self.inplanes, planes, stride, downsample))
314 |         self.inplanes = planes * block.expansion
315 |         for i in range(1, blocks):
316 |             layers.append(block(self.inplanes, planes))
317 | 
318 |         return nn.Sequential(*layers)
319 | 
320 |     def _make_up_conv_layer(self, up_proj_block, in_channels, out_channels):
321 |         return up_proj_block(in_channels, out_channels)
322 | 
323 |     def _make_gud_up_conv_layer(self, up_proj_block, in_channels, out_channels, oheight, owidth):
324 |         return up_proj_block(in_channels, out_channels, oheight, owidth)
325 | 
326 |     def _make_post_process_layer(self):
327 |         return post_process.Affinity_Propagate(spn=True)
328 | 
329 |     def forward(self, x):
330 |         [batch_size, channel, height, width] = x.size()
331 |         #sparse_depth = x.narrow(1, 3, 1).clone()  # get sparse depth
332 | 
333 |         x = self.conv1_1(x)
334 |         skip4 = x
335 | 
336 |         x = self.bn1(x)
337 |         x = self.relu(x)
338 |         x = self.maxpool(x)
339 |         x = self.layer1(x)
340 |         skip3 = x
341 | 
342 |         x = self.layer2(x)
343 |         skip2 = x
344 | 
345 |         x = self.layer3(x)
346 |         x = self.layer4(x)
347 | 
348 |         x = self.bn2(self.conv2(x))
349 |         x = self.gud_up_proj_layer1(x)
350 | 
351 |         x = self.gud_up_proj_layer2(x, skip2)
352 |         x = self.gud_up_proj_layer3(x, skip3)
353 |         x = self.gud_up_proj_layer4(x, skip4)
354 | 
355 |         blur_depth = self.gud_up_proj_layer5(x)
356 |         guidance = self.gud_up_proj_layer6(x)
357 |         #x = self.post_process_layer(guidance, blur_depth, sparse_depth)
358 |         
359 |         x = guidance
360 |         return x
361 | 
362 | 
363 | def resnet18(pretrained=False, **kwargs):
364 |     """Constructs a ResNet-18 model.
365 |     Args:
366 |         pretrained (bool): If True, returns a model pre-trained on ImageNet
367 |     """
368 |     model = ResNet(BasicBlock, [2, 2, 2, 2], UpProj_Block, **kwargs)
369 |     if pretrained:
370 |         print('==> Load pretrained model..')
371 |         pretrained_dict = torch.load(model_path['resnet18'])
372 |         model.load_state_dict(update_model.update_model(model, pretrained_dict))
373 |     return model
374 | 
375 | 
376 | def resnet34(pretrained=False, **kwargs):
377 |     """Constructs a ResNet-34 model.
378 |     Args:
379 |         pretrained (bool): If True, returns a model pre-trained on ImageNet
380 |     """
381 |     model = ResNet(BasicBlock, [3, 4, 6, 3], UpProj_Block, **kwargs)
382 |     if pretrained:
383 |         model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))
384 |     return model
385 | 
386 | 
387 | def resnet50(pretrained=False, **kwargs):
388 |     """Constructs a ResNet-50 model.
389 |     Args:
390 |         pretrained (bool): If True, returns a model pre-trained on ImageNet
391 |     """
392 |     model = ResNet(Bottleneck, [3, 4, 6, 3], UpProj_Block, **kwargs)
393 |     if pretrained:
394 |         print('==> Load pretrained model..')
395 |         pretrained_dict = torch.load(model_path['resnet50'])
396 |         model.load_state_dict(update_model.update_model(model, pretrained_dict))
397 |     return model
398 | 
399 | 
400 | def resnet101(pretrained=False, **kwargs):
401 |     """Constructs a ResNet-101 model.
402 |     Args:
403 |         pretrained (bool): If True, returns a model pre-trained on ImageNet
404 |     """
405 |     model = ResNet(Bottleneck, [3, 4, 23, 3], UpProj_Block, **kwargs)
406 |     if pretrained:
407 |         model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
408 |     return model
409 | 
410 | 
411 | def resnet152(pretrained=False, **kwargs):
412 |     """Constructs a ResNet-152 model.
413 |     Args:
414 |         pretrained (bool): If True, returns a model pre-trained on ImageNet
415 |     """
416 |     model = ResNet(Bottleneck, [3, 8, 36, 3], UpProj_Block, **kwargs)
417 |     if pretrained:
418 |         model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
419 |     return model
420 | 
421 | 
422 | if __name__ == "__main__":
423 |     image = torch.randn(1, 4, 228, 304).float()
424 |     image = image.cuda()
425 | 
426 |     model = resnet50(pretrained=True)
427 |     model = model.cuda()
428 | 
429 |     with torch.no_grad():
430 |         pred = model.forward(image)
431 | 
432 |     print(pred.size())
433 | 


--------------------------------------------------------------------------------
/nyu_dataloader.py:
--------------------------------------------------------------------------------
  1 | import os
  2 | import pickle
  3 | 
  4 | import numpy as np
  5 | import glob
  6 | import torch
  7 | from PIL import Image
  8 | from skimage.util import img_as_float
  9 | from skimage.transform import resize
 10 | from torchvision import transforms
 11 | 
 12 | 
 13 | class DataLoader():
 14 |     '''
 15 |     raw_data_dir = dir containing extracted raw NYU data (folder containing 2011-09-26 etc.)
 16 |     '''
 17 |     
 18 |     def __init__(self, raw_images_path, mode="train"):
 19 |         
 20 |         self.mode = mode
 21 |         self.raw_images_path = raw_images_path
 22 |         self.split_file = "nyu_split.pkl"
 23 |         
 24 |         if not os.path.isfile(self.split_file):
 25 |             self.train_test_split()
 26 |         
 27 |         with open(self.split_file, "rb") as split_f:
 28 |             split = pickle.load(split_f)
 29 |         
 30 |         if self.mode == "train":
 31 |             self.images = split["train"]
 32 |         else:
 33 |             self.images = split["val"]
 34 |         
 35 |         print('Found Images %d'%(len(self.images)))
 36 | 
 37 |     def load_data(self, img_file, label_file):
 38 |         x = Image.open(img_file) 
 39 |         y = Image.open(label_file)
 40 |         x = np.array(x)
 41 |         y = np.array(y)
 42 | 
 43 |         x = resize(x, (228, 304), anti_aliasing=True)
 44 |         y = resize(y, (228, 304), anti_aliasing=True)
 45 |         x = img_as_float(x)
 46 |         y = img_as_float(y)
 47 |         
 48 |         return x, y
 49 |     
 50 |     def __len__(self):
 51 |         return len(self.images)
 52 |     
 53 |     def train_test_split(self):
 54 |         images = [x.split('/')[-1].split('.')[0] for x in sorted(glob.glob(self.raw_images_path + '/*.jpg'))]
 55 | 
 56 |         total = len(images)
 57 |         indexes = np.random.permutation(total)
 58 | 
 59 |         split = int(0.9 * total)
 60 |         train = images[:split]
 61 |         val = images[split:]
 62 |         out = {'train': train, 'val': val}
 63 |         with open("nyu_split.pkl", "wb") as file:
 64 |             pickle.dump(out, file)
 65 |         
 66 |     def get_one_batch(self, batch_size = 16):
 67 |         images = []
 68 |         labels = []
 69 | 
 70 |         while True:
 71 |             idx = np.random.choice(len(self.images), batch_size)
 72 |             for i in idx:
 73 |                 image = os.path.join(self.raw_images_path, self.images[i] + ".jpg")
 74 |                 label = os.path.join(self.raw_images_path, self.images[i] + ".png")
 75 |                 x, y = self.load_data(image, label)
 76 |                 #x, y = self.train_transform(x, y)
 77 |                 y= np.expand_dims(y, axis=0)
 78 |                 images.append(x)
 79 |                 labels.append(y)
 80 |             yield torch.from_numpy(np.array(images)).permute(0, 3, 1, 2), torch.from_numpy(np.array(labels))
 81 | 
 82 |             
 83 |     def train_transform(self, rgb, depth):
 84 |         s = np.random.uniform(1.0, 1.5) # random scaling
 85 |         depth_np = depth / s
 86 |         angle = np.random.uniform(-5.0, 5.0) # random rotation degrees
 87 |         do_flip = np.random.uniform(0.0, 1.0) < 0.5 # random horizontal flip
 88 | 
 89 |         # perform 1st step of data augmentation
 90 |         transform = transforms.Compose([
 91 |             #transforms.Rotate(angle),
 92 |             transforms.Resize(s),
 93 |             transforms.CenterCrop(self.output_size),
 94 |             transforms.HorizontalFlip(do_flip)
 95 |         ])
 96 |         
 97 |         rgb_np = transform(rgb)
 98 |         rgb_np = self.color_jitter(rgb_np) # random color jittering
 99 |         rgb_np = np.asfarray(rgb_np, dtype='float') / 255
100 |         depth_np = transform(depth_np)
101 | 
102 |         return rgb_np, depth_np


--------------------------------------------------------------------------------
/nyu_loader.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import cv2
 3 | import glob
 4 | import torch
 5 | from PIL import Image
 6 | from skimage.util import img_as_float
 7 | from skimage.transform import resize
 8 | from torchvision import transforms
 9 | 
10 | class DataLoader():
11 |     '''
12 |     raw_data_dir = dir containing extracted raw NYU data (folder containing 2011-09-26 etc.)
13 |     '''
14 |     
15 |     def __init__(self, raw_images_path, mode="train"):
16 |         
17 |         self.mode = mode
18 |         
19 |         self.train_images = sorted(glob.glob(raw_images_path + '/*.jpg'))
20 |         self.train_labels = sorted(glob.glob(raw_images_path + '/*.png'))
21 |         
22 |         print('Found %d Images %d'%(len(self.train_images), len(self.train_labels)))
23 |     
24 |     def load_data(self, img_file, label_file):
25 |         x = Image.open(img_file) 
26 |         y = Image.open(label_file).convert('RGB')
27 |         x = np.array(x).astype(np.float32)
28 |         y = np.array(y).astype(np.float32)
29 | 
30 |         x = resize(x, (228, 304), anti_aliasing=True)
31 |         y = resize(y, (228, 304), anti_aliasing=True)
32 |         x = img_as_float(x)/127.5 - 1
33 |         y = img_as_float(y)/127.5 - 1
34 |         
35 |         return x, y
36 |         
37 |     def get_one_batch(self, batch_size = 16):
38 |         images = []
39 |         labels = []
40 | 
41 |         while True:
42 |             idx = np.random.choice(len(self.train_images), batch_size)
43 |             for i in idx:
44 |                 x, y = self.load_data(self.train_images[i], self.train_labels[i])
45 |                 #x, y = self.train_transform(x, y)
46 |                 images.append(x)
47 |                 labels.append(y)
48 |             yield torch.from_numpy(np.array(images)).permute(0, 3, 1, 2), torch.from_numpy(np.array(labels)).permute(0, 3, 1, 2)
49 | 
50 |             
51 |     def train_transform(self, rgb, depth):
52 |         s = np.random.uniform(1.0, 1.5) # random scaling
53 |         depth_np = depth / s
54 |         angle = np.random.uniform(-5.0, 5.0) # random rotation degrees
55 |         do_flip = np.random.uniform(0.0, 1.0) < 0.5 # random horizontal flip
56 | 
57 |         # perform 1st step of data augmentation
58 |         transform = transforms.Compose([
59 |             #transforms.Rotate(angle),
60 |             transforms.Resize(s),
61 |             transforms.CenterCrop(self.output_size),
62 |             transforms.HorizontalFlip(do_flip)
63 |         ])
64 |         
65 |         rgb_np = transform(rgb)
66 |         rgb_np = self.color_jitter(rgb_np) # random color jittering
67 |         rgb_np = np.asfarray(rgb_np, dtype='float') / 255
68 |         depth_np = transform(depth_np)
69 | 
70 |         return rgb_np, depth_np
71 | 
72 | 


--------------------------------------------------------------------------------
/nyu_train.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import cv2
  3 | import torch
  4 | import torch.nn as nn
  5 | import torch.nn.functional as F
  6 | import torchvision.models
  7 | import collections
  8 | import math
  9 | import time
 10 | import pickle
 11 | 
 12 | from nyu_loader import DataLoader
 13 | 
 14 | def weights_init(m):
 15 |     # Initialize filters with Gaussian random weights
 16 |     if isinstance(m, nn.Conv2d):
 17 |         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
 18 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 19 |         if m.bias is not None:
 20 |             m.bias.data.zero_()
 21 |     elif isinstance(m, nn.ConvTranspose2d):
 22 |         n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
 23 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 24 |         if m.bias is not None:
 25 |             m.bias.data.zero_()
 26 |     elif isinstance(m, nn.BatchNorm2d):
 27 |         m.weight.data.fill_(1)
 28 |         m.bias.data.zero_()
 29 | 
 30 | 
 31 | class Unpool(nn.Module):
 32 |     # Unpool: 2*2 unpooling with zero padding
 33 |     def __init__(self, num_channels, stride=2):
 34 |         super(Unpool, self).__init__()
 35 | 
 36 |         self.num_channels = num_channels
 37 |         self.stride = stride
 38 | 
 39 |     def forward(self, x):
 40 |         weights = torch.zeros(self.num_channels, 1, self.stride, self.stride)
 41 |         if torch.cuda.is_available():
 42 |             weights = weights.cuda()
 43 |         weights[:, :, 0, 0] = 1
 44 |         return F.conv_transpose2d(x, weights, stride=self.stride, groups=self.num_channels)
 45 | 
 46 | 
 47 | class Decoder(nn.Module):
 48 |     names = ['deconv2', 'deconv3', 'upconv', 'upproj']
 49 | 
 50 |     def __init__(self):
 51 |         super(Decoder, self).__init__()
 52 | 
 53 |         self.layer1 = None
 54 |         self.layer2 = None
 55 |         self.layer3 = None
 56 |         self.layer4 = None
 57 | 
 58 |     def forward(self, x):
 59 |         x = self.layer1(x)
 60 |         x = self.layer2(x)
 61 |         x = self.layer3(x)
 62 |         x = self.layer4(x)
 63 |         return x
 64 | 
 65 | 
 66 | class DeConv(Decoder):
 67 |     def __init__(self, in_channels, kernel_size):
 68 |         assert kernel_size >= 2, "kernel_size out of range: {}".format(kernel_size)
 69 |         super(DeConv, self).__init__()
 70 | 
 71 |         def convt(in_channels):
 72 |             stride = 2
 73 |             padding = (kernel_size - 1) // 2
 74 |             output_padding = kernel_size % 2
 75 |             assert -2 - 2 * padding + kernel_size + output_padding == 0, "deconv parameters incorrect"
 76 | 
 77 |             module_name = "deconv{}".format(kernel_size)
 78 |             return nn.Sequential(collections.OrderedDict([
 79 |                 (module_name, nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size,
 80 |                                                  stride, padding, output_padding, bias=False)),
 81 |                 ('batchnorm', nn.BatchNorm2d(in_channels // 2)),
 82 |                 ('relu', nn.ReLU(inplace=True)),
 83 |             ]))
 84 | 
 85 |         self.layer1 = convt(in_channels)
 86 |         self.layer2 = convt(in_channels // 2)
 87 |         self.layer3 = convt(in_channels // (2 ** 2))
 88 |         self.layer4 = convt(in_channels // (2 ** 3))
 89 | 
 90 | 
 91 | class UpConv(Decoder):
 92 |     # UpConv decoder consists of 4 upconv modules with decreasing number of channels and increasing feature map size
 93 |     def upconv_module(self, in_channels):
 94 |         # UpConv module: unpool -> 5*5 conv -> batchnorm -> ReLU
 95 |         upconv = nn.Sequential(collections.OrderedDict([
 96 |             ('unpool', Unpool(in_channels)),
 97 |             ('conv', nn.Conv2d(in_channels, in_channels // 2, kernel_size=5, stride=1, padding=2, bias=False)),
 98 |             ('batchnorm', nn.BatchNorm2d(in_channels // 2)),
 99 |             ('relu', nn.ReLU()),
100 |         ]))
101 |         return upconv
102 | 
103 |     def __init__(self, in_channels):
104 |         super(UpConv, self).__init__()
105 |         self.layer1 = self.upconv_module(in_channels)
106 |         self.layer2 = self.upconv_module(in_channels // 2)
107 |         self.layer3 = self.upconv_module(in_channels // 4)
108 |         self.layer4 = self.upconv_module(in_channels // 8)
109 | 
110 | 
111 | class FasterUpConv(Decoder):
112 |     # Faster Upconv using pixelshuffle
113 | 
114 |     class faster_upconv_module(nn.Module):
115 | 
116 |         def __init__(self, in_channel):
117 |             super(FasterUpConv.faster_upconv_module, self).__init__()
118 | 
119 |             self.conv1_ = nn.Sequential(collections.OrderedDict([
120 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=3)),
121 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
122 |             ]))
123 | 
124 |             self.conv2_ = nn.Sequential(collections.OrderedDict([
125 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(2, 3))),
126 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
127 |             ]))
128 | 
129 |             self.conv3_ = nn.Sequential(collections.OrderedDict([
130 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(3, 2))),
131 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
132 |             ]))
133 | 
134 |             self.conv4_ = nn.Sequential(collections.OrderedDict([
135 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=2)),
136 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
137 |             ]))
138 | 
139 |             self.ps = nn.PixelShuffle(2)
140 |             self.relu = nn.ReLU(inplace=True)
141 | 
142 |         def forward(self, x):
143 |             # print('Upmodule x size = ', x.size())
144 |             x1 = self.conv1_(nn.functional.pad(x, (1, 1, 1, 1)))
145 |             x2 = self.conv2_(nn.functional.pad(x, (1, 1, 0, 1)))
146 |             x3 = self.conv3_(nn.functional.pad(x, (0, 1, 1, 1)))
147 |             x4 = self.conv4_(nn.functional.pad(x, (0, 1, 0, 1)))
148 | 
149 |             x = torch.cat((x1, x2, x3, x4), dim=1)
150 | 
151 |             output = self.ps(x)
152 |             output = self.relu(output)
153 | 
154 |             return output
155 | 
156 |     def __init__(self, in_channel):
157 |         super(FasterUpConv, self).__init__()
158 | 
159 |         self.layer1 = self.faster_upconv_module(in_channel)
160 |         self.layer2 = self.faster_upconv_module(in_channel // 2)
161 |         self.layer3 = self.faster_upconv_module(in_channel // 4)
162 |         self.layer4 = self.faster_upconv_module(in_channel // 8)
163 | 
164 | 
165 | class UpProj(Decoder):
166 |     # UpProj decoder consists of 4 upproj modules with decreasing number of channels and increasing feature map size
167 | 
168 |     class UpProjModule(nn.Module):
169 |         # UpProj module has two branches, with a Unpool at the start and a ReLu at the end
170 |         #   upper branch: 5*5 conv -> batchnorm -> ReLU -> 3*3 conv -> batchnorm
171 |         #   bottom branch: 5*5 conv -> batchnorm
172 | 
173 |         def __init__(self, in_channels):
174 |             super(UpProj.UpProjModule, self).__init__()
175 |             out_channels = in_channels // 2
176 |             self.unpool = Unpool(in_channels)
177 |             self.upper_branch = nn.Sequential(collections.OrderedDict([
178 |                 ('conv1', nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)),
179 |                 ('batchnorm1', nn.BatchNorm2d(out_channels)),
180 |                 ('relu', nn.ReLU()),
181 |                 ('conv2', nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)),
182 |                 ('batchnorm2', nn.BatchNorm2d(out_channels)),
183 |             ]))
184 |             self.bottom_branch = nn.Sequential(collections.OrderedDict([
185 |                 ('conv', nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)),
186 |                 ('batchnorm', nn.BatchNorm2d(out_channels)),
187 |             ]))
188 |             self.relu = nn.ReLU()
189 | 
190 |         def forward(self, x):
191 |             x = self.unpool(x)
192 |             x1 = self.upper_branch(x)
193 |             x2 = self.bottom_branch(x)
194 |             x = x1 + x2
195 |             x = self.relu(x)
196 |             return x
197 | 
198 |     def __init__(self, in_channels):
199 |         super(UpProj, self).__init__()
200 |         self.layer1 = self.UpProjModule(in_channels)
201 |         self.layer2 = self.UpProjModule(in_channels // 2)
202 |         self.layer3 = self.UpProjModule(in_channels // 4)
203 |         self.layer4 = self.UpProjModule(in_channels // 8)
204 | 
205 | 
206 | class FasterUpProj(Decoder):
207 |     # Faster UpProj decorder using pixelshuffle
208 | 
209 |     class faster_upconv(nn.Module):
210 | 
211 |         def __init__(self, in_channel):
212 |             super(FasterUpProj.faster_upconv, self).__init__()
213 | 
214 |             self.conv1_ = nn.Sequential(collections.OrderedDict([
215 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=3)),
216 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
217 |             ]))
218 | 
219 |             self.conv2_ = nn.Sequential(collections.OrderedDict([
220 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(2, 3))),
221 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
222 |             ]))
223 | 
224 |             self.conv3_ = nn.Sequential(collections.OrderedDict([
225 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(3, 2))),
226 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
227 |             ]))
228 | 
229 |             self.conv4_ = nn.Sequential(collections.OrderedDict([
230 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=2)),
231 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
232 |             ]))
233 | 
234 |             self.ps = nn.PixelShuffle(2)
235 |             self.relu = nn.ReLU(inplace=True)
236 | 
237 |         def forward(self, x):
238 |             # print('Upmodule x size = ', x.size())
239 |             x1 = self.conv1_(nn.functional.pad(x, (1, 1, 1, 1)))
240 |             x2 = self.conv2_(nn.functional.pad(x, (1, 1, 0, 1)))
241 |             x3 = self.conv3_(nn.functional.pad(x, (0, 1, 1, 1)))
242 |             x4 = self.conv4_(nn.functional.pad(x, (0, 1, 0, 1)))
243 |             # print(x1.size(), x2.size(), x3.size(), x4.size())
244 | 
245 |             x = torch.cat((x1, x2, x3, x4), dim=1)
246 | 
247 |             x = self.ps(x)
248 |             return x
249 | 
250 |     class FasterUpProjModule(nn.Module):
251 |         def __init__(self, in_channels):
252 |             super(FasterUpProj.FasterUpProjModule, self).__init__()
253 |             out_channels = in_channels // 2
254 | 
255 |             self.upper_branch = nn.Sequential(collections.OrderedDict([
256 |                 ('faster_upconv', FasterUpProj.faster_upconv(in_channels)),
257 |                 ('relu', nn.ReLU(inplace=True)),
258 |                 ('conv', nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)),
259 |                 ('batchnorm', nn.BatchNorm2d(out_channels)),
260 |             ]))
261 |             self.bottom_branch = FasterUpProj.faster_upconv(in_channels)
262 |             self.relu = nn.ReLU(inplace=True)
263 | 
264 |         def forward(self, x):
265 |             x1 = self.upper_branch(x)
266 |             x2 = self.bottom_branch(x)
267 |             x = x1 + x2
268 |             x = self.relu(x)
269 |             return x
270 | 
271 |     def __init__(self, in_channel):
272 |         super(FasterUpProj, self).__init__()
273 | 
274 |         self.layer1 = self.FasterUpProjModule(in_channel)
275 |         self.layer2 = self.FasterUpProjModule(in_channel // 2)
276 |         self.layer3 = self.FasterUpProjModule(in_channel // 4)
277 |         self.layer4 = self.FasterUpProjModule(in_channel // 8)
278 | 
279 | 
280 | def choose_decoder(decoder, in_channels):
281 |     if decoder[:6] == 'deconv':
282 |         assert len(decoder) == 7
283 |         kernel_size = int(decoder[6])
284 |         return DeConv(in_channels, kernel_size)
285 |     elif decoder == "upproj":
286 |         return UpProj(in_channels)
287 |     elif decoder == "upconv":
288 |         return UpConv(in_channels)
289 |     elif decoder == "fasterupproj":
290 |         return FasterUpProj(in_channels)
291 |     else:
292 |         assert False, "invalid option for decoder: {}".format(decoder)
293 | 
294 | class ResNet(nn.Module):
295 |     def __init__(self, dataset = 'kitti', layers = 50, decoder = 'upproj', output_size=(228, 304), in_channels=3, pretrained=True):
296 | 
297 |         if layers not in [18, 34, 50, 101, 152]:
298 |             raise RuntimeError(
299 |                 'Only 18, 34, 50, 101, and 152 layer model are defined for ResNet. Got {}'.format(layers))
300 | 
301 |         super(ResNet, self).__init__()
302 |         pretrained_model = torchvision.models.__dict__['resnet{}'.format(layers)](pretrained=pretrained)
303 | 
304 |         if in_channels == 3:
305 |             self.conv1 = pretrained_model._modules['conv1']
306 |             self.bn1 = pretrained_model._modules['bn1']
307 |         else:
308 |             self.conv1 = nn.Conv2d(in_channels, 64, kernel_size=7, stride=2, padding=3, bias=False)
309 |             self.bn1 = nn.BatchNorm2d(64)
310 |             weights_init(self.conv1)
311 |             weights_init(self.bn1)
312 | 
313 |         self.output_size = output_size
314 | 
315 |         self.relu = pretrained_model._modules['relu']
316 |         self.maxpool = pretrained_model._modules['maxpool']
317 |         self.layer1 = pretrained_model._modules['layer1']
318 |         self.layer2 = pretrained_model._modules['layer2']
319 |         self.layer3 = pretrained_model._modules['layer3']
320 |         self.layer4 = pretrained_model._modules['layer4']
321 | 
322 |         # clear memory
323 |         del pretrained_model
324 | 
325 |         # define number of intermediate channels
326 |         if layers <= 34:
327 |             num_channels = 512
328 |         elif layers >= 50:
329 |             num_channels = 2048
330 | 
331 |         self.conv2 = nn.Conv2d(num_channels, num_channels // 2, kernel_size=1, bias=False)
332 |         self.bn2 = nn.BatchNorm2d(num_channels // 2)
333 | 
334 |         self.upSample = choose_decoder(decoder, num_channels // 2)
335 | 
336 |         # setting bias=true doesn't improve accuracy
337 |         self.conv3 = nn.Conv2d(num_channels // 32, 3, kernel_size=3, stride=1, padding=1, bias=False)
338 |         self.bilinear = nn.Upsample(size=self.output_size, mode='bilinear', align_corners=True)
339 | 
340 |         # weight init
341 |         self.conv2.apply(weights_init)
342 |         self.bn2.apply(weights_init)
343 | 
344 |         self.upSample.apply(weights_init)
345 | 
346 |         self.conv3.apply(weights_init)
347 | 
348 |     def forward(self, x):
349 |         # resnet
350 |         x = self.conv1(x)
351 |         x = self.bn1(x)
352 |         x = self.relu(x)
353 |         x = self.maxpool(x)
354 |         x1 = self.layer1(x)
355 |         x2 = self.layer2(x1)
356 |         x3 = self.layer3(x2)
357 |         x4 = self.layer4(x3)
358 | 
359 |         x = self.conv2(x4)
360 |         x = self.bn2(x)
361 | 
362 |         x = self.upSample(x)
363 | 
364 |         x = self.conv3(x)
365 |         x = self.bilinear(x)
366 | 
367 |         return x
368 | 
369 |     def get_1x_lr_params(self):
370 |         """
371 |         This generator returns all the parameters of the net layer whose learning rate is 1x lr.
372 |         """
373 |         b = [self.conv1, self.bn1, self.relu, self.maxpool, self.layer1, self.layer2, self.layer3, self.layer4]
374 |         for i in range(len(b)):
375 |             for k in b[i].parameters():
376 |                 if k.requires_grad:
377 |                     yield k
378 | 
379 |     def get_10x_lr_params(self):
380 |         """
381 |         This generator returns all the parameters of the net layer whose learning rate is 20x lr.
382 |         """
383 |         b = [self.conv2, self.bn2, self.upSample, self.conv3, self.bilinear]
384 |         for j in range(len(b)):
385 |             for k in b[j].parameters():
386 |                 if k.requires_grad:
387 |                     yield k
388 | 
389 | 
390 | class Discriminator(nn.Module):
391 |     """
392 |     Discriminator for FCRN
393 |     """
394 |     def __init__(self, n_in):
395 |         super(Discriminator, self).__init__()
396 |         self.n_in = n_in
397 |         self.modules = []
398 | 
399 |         self.modules += [   nn.Conv2d(self.n_in, 32, kernel_size=3, stride=(1, 1), padding=(2, 2)),
400 |                             nn.MaxPool2d(kernel_size=(2, 2)),
401 |                             nn.LeakyReLU(0.2, True),
402 |                             nn.Conv2d(32, 32, 3, 1, 1),
403 |                             nn.MaxPool2d(kernel_size=(2,2)),
404 |                             nn.LeakyReLU(0.2, True),
405 |                             nn.Conv2d(32, 32, 3, 1, 1),
406 |                             nn.MaxPool2d(kernel_size=(2,2)),
407 |                             nn.LeakyReLU(0.2, True),
408 |                             nn.Conv2d(32, 32, 3, 1, 1),
409 |                             nn.MaxPool2d(kernel_size=(2,2)),
410 |                             nn.LeakyReLU(0.2, True),
411 |         ]
412 | 
413 |         self.model = nn.Sequential(*self.modules)
414 |         self.fc = nn.Linear(32 * 4, 1)
415 |         
416 |     def forward(self, x):
417 |         x = self.model(x)
418 |         x = x.view(x.size(0), -1)
419 |         x = self.fc(x)
420 |         return torch.sigmoid(x)
421 | 
422 | 
423 | """
424 | Define Loss functions here
425 | """
426 | class ScaleInvariantError(nn.Module):
427 |     """
428 |     Scale invariant error defined in Eigen's paper!
429 |     """
430 | 
431 |     def __init__(self, lamada=1.0):
432 |         super(ScaleInvariantError, self).__init__()
433 |         self.lamada = lamada
434 |         return
435 | 
436 |     def forward(self, y_true, y_pred):
437 |         first_log = torch.log(torch.clamp(y_pred, 0, 1))
438 |         second_log = torch.log(torch.clamp(y_true, 0, 1))
439 |         d = first_log - second_log
440 |         loss = torch.mean(d * d) - self.lamada * torch.mean(d) * torch.mean(d)
441 |         return loss
442 | 
443 | 
444 | class MaskedMSELoss(nn.Module):
445 |     def __init__(self):
446 |         super(MaskedMSELoss, self).__init__()
447 | 
448 |     def forward(self, pred, target):
449 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
450 |         valid_mask = (target > 0).detach()
451 |         diff = target - pred
452 |         diff = diff[valid_mask]
453 |         self.loss = (diff ** 2).mean()
454 |         return self.loss
455 | 
456 | 
457 | class MaskedL1Loss(nn.Module):
458 |     def __init__(self):
459 |         super(MaskedL1Loss, self).__init__()
460 | 
461 |     def forward(self, pred, target):
462 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
463 |         valid_mask = (target > 0).detach()
464 |         diff = target - pred
465 |         diff = diff[valid_mask]
466 |         self.loss = diff.abs().mean()
467 |         return self.loss
468 | 
469 | 
470 | class berHuLoss(nn.Module):
471 |     def __init__(self):
472 |         super(berHuLoss, self).__init__()
473 | 
474 |     def forward(self, pred, target):
475 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
476 | 
477 |         huber_c = torch.max(pred - target)
478 |         huber_c = 0.2 * huber_c
479 | 
480 |         valid_mask = (target > 0).detach()
481 |         diff = target - pred
482 |         diff = diff[valid_mask]
483 |         diff = diff.abs()
484 | 
485 |         huber_mask = (diff > huber_c).detach()
486 | 
487 |         diff2 = diff[huber_mask]
488 |         diff2 = diff2 ** 2
489 | 
490 |         self.loss = torch.cat((diff, diff2)).mean()
491 | 
492 |         return self.loss
493 | 
494 |     
495 | def set_requires_grad(net, requires_grad=False):
496 |     if net is not None:
497 |         for param in net.parameters():
498 |             param.requires_grad = requires_grad
499 | 
500 | def save_image(model, x, batch, mode="train"):
501 |     pred = model(x)
502 |     pred *= 100
503 |     npimg = pred.cpu().detach().numpy()
504 |     npimg = np.transpose(npimg, (0, 2, 3, 1))
505 |     cv2.imwrite('saved_images/%s_image_%d.jpg'%(mode, batch), npimg[0])
506 | 
507 | 
508 | def train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, 
509 |           criterion_berHu, criterion_GAN, optimizer, optimizer_D, epoch, batch_size, val_frequency=10):
510 |     model.train()  # switch to train mode
511 |     eval_mode = False
512 |     num_batches = 1453 // batch_size
513 |     init_lr = optimizer.param_groups[0]['lr']
514 |     
515 |     total_losses_log = {"l1":0, "mse":0, "berHu":0, "adv":0}
516 |     
517 |     valid_T = torch.ones(batch_size, 1).cuda()
518 |     zeros_T = torch.zeros(batch_size, 1).cuda()
519 |     
520 |     train_loss_list = []
521 |     val_loss_list = []
522 |     
523 |     for iter_ in range(num_batches):
524 |         # Adjust Learning Rate
525 |         if iter_ % 1000 == 0 and optimizer.param_groups[0]['lr'] > init_lr/10.0:
526 |             for param_group in optimizer.param_groups:
527 |                 param_group['lr'] *= 0.98
528 |         
529 |         input, target = next(train_loader.get_one_batch(batch_size))
530 |         input, target = input.float(), target.float()
531 |         input, target = input.cuda(), target.cuda()
532 |         torch.cuda.synchronize()
533 |         
534 |         optimizer.zero_grad()
535 |         
536 |         pred = model(input)
537 |         
538 |         loss_L1 = criterion_L1(pred, target)
539 |         loss_MSE = criterion_MSE(pred, target)
540 |         loss_berHu = criterion_berHu(pred, target)
541 |         #loss_SI = criterion_SI(pred, target)
542 |         '''
543 |         set_requires_grad(discriminator, False)
544 |         
545 |         loss_adv = 0
546 |         
547 |         for a in range(3):
548 |             for b in range(9):
549 |                 row = 30 * a
550 |                 col = 30 * b
551 |                 patch_fake = pred[:, :, row:row+30, col:col+30]
552 |                 pred_fake = discriminator(patch_fake)
553 |                 loss_adv += criterion_GAN(pred_fake, valid_T)
554 |         '''
555 |         
556 |         loss_gen = loss_L1 + loss_berHu
557 |         loss_gen.backward()
558 |         optimizer.step()
559 |         '''
560 |         set_requires_grad(discriminator, True)
561 |         optimizer_D.zero_grad()
562 |         loss_D = 0
563 |         for a in range(3):
564 |             for b in range(9):
565 |                 row = 30 * a
566 |                 col = 30 * b
567 |                 patch_fake = pred[:, :, row:row+30, col:col+30]
568 |                 patch_real = target[:, :, row:row+30, col:col+30]
569 |                 pred_fake = discriminator(patch_fake.detach())
570 |                 pred_real = discriminator(patch_real)
571 |                 loss_D_fake = criterion_GAN(pred_fake, zeros_T)
572 |                 loss_D_real = criterion_GAN(pred_real, valid_T)
573 |                 loss_D += 0.5 * (loss_D_fake + loss_D_real)
574 |         
575 |         loss_D.backward()
576 |         optimizer_D.step()
577 | 
578 |         torch.cuda.synchronize()
579 |         '''
580 |         total_losses_log["l1"] += loss_L1.item()
581 |         total_losses_log["mse"] += loss_MSE.item()
582 |         total_losses_log["berHu"] += loss_berHu.item()
583 |         #total_losses_log["adv"] += loss_adv.item()
584 |          
585 |         if (iter_ + 1) % 10 == 0:
586 |             save_image(model, input, iter_)
587 |             print('Train Epoch: {} Batch: [{}/{}], L1 ={:0.3f}, MSE={:0.3f}, berHu={:0.3f}'.format(
588 |                 epoch, iter_ + 1, num_batches, 
589 |                 loss_L1.item(), loss_MSE.item(), loss_berHu.item()))
590 |         
591 |         if (iter_ + 1) % val_frequency == 0:
592 |             val_loss = validate(val_loader, model, discriminator, criterion_L1, criterion_MSE, criterion_berHu, criterion_GAN, batch_size, iter_)
593 |             for key in total_losses_log.keys():
594 |                 total_losses_log[key] /= (val_frequency * batch_size)
595 |             train_loss_list.append(total_losses_log)
596 |             val_loss_list.append(val_loss)
597 |             save_losses(train_loss_list, val_loss_list)
598 |             print("TRAIN:- ", total_losses_log)
599 |             print("VAL :- ", val_loss)
600 |             total_losses_log = {"l1":0, "mse":0, "berHu":0, "adv":0}
601 | 
602 | def save_losses(train_loss_list, val_loss_list):
603 |     out = {"train_loss": train_loss_list, "val_loss":val_loss_list}
604 |     with open("losses.pkl", "wb") as log:
605 |         pickle.dump(out, log)
606 |         
607 | def validate(val_loader, model, discriminator, criterion_L1, criterion_MSE, 
608 |              criterion_berHu, criterion_GAN, batch_size, train_iter):
609 |     model.eval()  # switch to evaluate mode
610 |     #discriminator.eval()
611 |     
612 |     valid_T = torch.ones(batch_size, 1).cuda()
613 |     zeros_T = torch.zeros(batch_size, 1).cuda()
614 |     
615 |     #num_batches = len(val_loader) // batch_size
616 |     total_losses = {"l1":0, "mse":0, "berHu":0, "adv":0}
617 |     num_batches = 25
618 |     for i in range(num_batches):
619 | 
620 |         input, target = next(val_loader.get_one_batch(batch_size))
621 |         input, target = input.float(), target.float()
622 |         input, target = input.cuda(), target.cuda()
623 |         
624 |         torch.cuda.synchronize()
625 | 
626 |         # compute output
627 |         end = time.time()
628 |         with torch.no_grad():
629 |             pred = model(input)
630 |             
631 |         torch.cuda.synchronize()
632 |         
633 |         loss_L1 = criterion_L1(pred, target)
634 |         loss_MSE = criterion_MSE(pred, target)
635 |         loss_berHu = criterion_berHu(pred, target)
636 |         '''
637 |         loss_adv = 0
638 |         with torch.no_grad():
639 |             for a in range(3):
640 |                 for b in range(9):
641 |                     row = 30 * a
642 |                     col = 30 * b
643 |                     patch_fake = pred[:, :, row:row+30, col:col+30]
644 |                     pred_fake = discriminator(patch_fake)
645 |                     loss_adv += criterion_GAN(pred_fake, valid_T)
646 |         '''            
647 |         total_losses["l1"] += loss_L1.item()
648 |         total_losses["mse"] += loss_MSE.item()
649 |         total_losses["berHu"] += loss_berHu.item()
650 |         #total_losses["adv"] += loss_adv.item()
651 |         
652 |         if i==0:
653 |             save_image(model, input, train_iter, "val")
654 |     
655 |     for key in total_losses.keys():
656 |         total_losses[key] /= (num_batches * batch_size)
657 |     
658 |     model.train()
659 |     #discriminator.train() 
660 |     return total_losses
661 | 
662 | raw_data_dir = ""
663 | depth_maps_dir = "depth/"
664 | nyu_dir = "../cnn_depth_tensorflow/data/nyu_datasets/"
665 | print("=> Loading Data ...")
666 | train_loader = DataLoader(nyu_dir)
667 | val_loader = DataLoader(nyu_dir)
668 | 
669 | print("=> creating Model")
670 | model = ResNet(layers=101, output_size=(90, 270), pretrained=True)
671 | discriminator = Discriminator(3)
672 | discriminator.apply(weights_init)
673 | 
674 | print("=> model created.")
675 | start_epoch = 0
676 | init_lr = 0.001
677 | 
678 | train_params = [{'params': model.get_1x_lr_params(), 'lr': init_lr},
679 |                 {'params': model.get_10x_lr_params(), 'lr': init_lr * 10}]
680 | 
681 | optimizer = torch.optim.SGD(train_params, lr=init_lr)
682 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=init_lr)
683 | 
684 | # You can use DataParallel() whether you use Multi-GPUs or not
685 | model = nn.DataParallel(model).cuda()
686 | #model = model.cuda()
687 | discriminator = discriminator.cuda()
688 | 
689 | # Define Loss Function
690 | criterion_L1 = MaskedL1Loss()
691 | criterion_berHu = berHuLoss()
692 | criterion_MSE = MaskedMSELoss()
693 | criterion_GAN = nn.BCELoss()
694 | 
695 | batch_size = 64
696 | val_frequency = 50
697 | 
698 | for epoch in range(50):
699 |     # Train the Model
700 |     train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, criterion_berHu, 
701 |           criterion_GAN, optimizer, optimizer_D, epoch, batch_size, val_frequency)
702 |      
703 |     # Save Checkpoint
704 |     torch.save({
705 |             'epoch': epoch,
706 |             'model_state_dict': model.state_dict(),
707 |             'optimizer_state_dict': optimizer.state_dict(),
708 |             }, './ResNet.pth')
709 | 
710 | 


--------------------------------------------------------------------------------
/prepare_data.py:
--------------------------------------------------------------------------------
 1 | #encoding: utf-8
 2 | import os
 3 | import numpy as np
 4 | import h5py
 5 | from PIL import Image
 6 | import random
 7 | import wget
 8 | 
 9 | 
10 | def convert_nyu(path):
11 |     imgdir = os.path.join("data", "nyu_datasets");
12 |     if not os.path.exists(imgdir):
13 |         os.makedirs(imgdir)
14 | 
15 |     nyuurl = 'http://horatio.cs.nyu.edu/mit/silberman/nyu_depth_v2/nyu_depth_v2_labeled.mat'
16 |     file = os.path.join("data", "nyu_depth_v2_labeled.mat")
17 |     if not os.path.exists(file):    
18 |         filename = wget.download(nyuurl, out="data")
19 |         print('\n downloaded: ', filename)
20 | 
21 |     print("load dataset: %s" % (path))
22 |     f = h5py.File(path)
23 | 
24 |     trains = []
25 |     for i, (image, depth) in enumerate(zip(f['images'], f['depths'])):
26 |         ra_image = image.transpose(2, 1, 0)
27 |         ra_depth = depth.transpose(1, 0)
28 |         re_depth = (ra_depth/np.max(ra_depth))*255.0
29 |         image_pil = Image.fromarray(np.uint8(ra_image))
30 |         depth_pil = Image.fromarray(np.uint8(re_depth))
31 |         image_name = os.path.join("data", "nyu_datasets", "%05d.jpg" % (i))
32 |         image_pil.save(image_name)
33 |         depth_name = os.path.join("data", "nyu_datasets", "%05d.png" % (i))
34 |         depth_pil.save(depth_name)
35 | 
36 |         trains.append((image_name, depth_name))
37 | 
38 |     random.shuffle(trains)
39 | 
40 |     if not os.path.exists('train.csv'):  
41 |         os.remove('train.csv')
42 | 
43 |     with open('train.csv', 'w') as output:
44 |         for (image_name, depth_name) in trains:
45 |             output.write("%s,%s" % (image_name, depth_name))
46 |             output.write("\n")
47 | 
48 | if __name__ == '__main__':
49 |     current_directory = os.getcwd()
50 |     nyu_path = 'data/nyu_depth_v2_labeled.mat'
51 |     convert_nyu(nyu_path)
52 | 


--------------------------------------------------------------------------------
/train.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import cv2
  3 | import torch
  4 | import torch.nn as nn
  5 | import torch.nn.functional as F
  6 | import torchvision.models
  7 | import collections
  8 | import math
  9 | import time
 10 | import pickle
 11 | 
 12 | from data_loader import DataLoader
 13 | 
 14 | def weights_init(m):
 15 |     # Initialize filters with Gaussian random weights
 16 |     if isinstance(m, nn.Conv2d):
 17 |         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
 18 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 19 |         if m.bias is not None:
 20 |             m.bias.data.zero_()
 21 |     elif isinstance(m, nn.ConvTranspose2d):
 22 |         n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
 23 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 24 |         if m.bias is not None:
 25 |             m.bias.data.zero_()
 26 |     elif isinstance(m, nn.BatchNorm2d):
 27 |         m.weight.data.fill_(1)
 28 |         m.bias.data.zero_()
 29 | 
 30 | 
 31 | class Unpool(nn.Module):
 32 |     # Unpool: 2*2 unpooling with zero padding
 33 |     def __init__(self, num_channels, stride=2):
 34 |         super(Unpool, self).__init__()
 35 | 
 36 |         self.num_channels = num_channels
 37 |         self.stride = stride
 38 | 
 39 |     def forward(self, x):
 40 |         weights = torch.zeros(self.num_channels, 1, self.stride, self.stride)
 41 |         if torch.cuda.is_available():
 42 |             weights = weights.cuda()
 43 |         weights[:, :, 0, 0] = 1
 44 |         return F.conv_transpose2d(x, weights, stride=self.stride, groups=self.num_channels)
 45 | 
 46 | 
 47 | class Decoder(nn.Module):
 48 |     names = ['deconv2', 'deconv3', 'upconv', 'upproj']
 49 | 
 50 |     def __init__(self):
 51 |         super(Decoder, self).__init__()
 52 | 
 53 |         self.layer1 = None
 54 |         self.layer2 = None
 55 |         self.layer3 = None
 56 |         self.layer4 = None
 57 | 
 58 |     def forward(self, x):
 59 |         x = self.layer1(x)
 60 |         x = self.layer2(x)
 61 |         x = self.layer3(x)
 62 |         x = self.layer4(x)
 63 |         return x
 64 | 
 65 | 
 66 | class DeConv(Decoder):
 67 |     def __init__(self, in_channels, kernel_size):
 68 |         assert kernel_size >= 2, "kernel_size out of range: {}".format(kernel_size)
 69 |         super(DeConv, self).__init__()
 70 | 
 71 |         def convt(in_channels):
 72 |             stride = 2
 73 |             padding = (kernel_size - 1) // 2
 74 |             output_padding = kernel_size % 2
 75 |             assert -2 - 2 * padding + kernel_size + output_padding == 0, "deconv parameters incorrect"
 76 | 
 77 |             module_name = "deconv{}".format(kernel_size)
 78 |             return nn.Sequential(collections.OrderedDict([
 79 |                 (module_name, nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size,
 80 |                                                  stride, padding, output_padding, bias=False)),
 81 |                 ('batchnorm', nn.BatchNorm2d(in_channels // 2)),
 82 |                 ('relu', nn.ReLU(inplace=True)),
 83 |             ]))
 84 | 
 85 |         self.layer1 = convt(in_channels)
 86 |         self.layer2 = convt(in_channels // 2)
 87 |         self.layer3 = convt(in_channels // (2 ** 2))
 88 |         self.layer4 = convt(in_channels // (2 ** 3))
 89 | 
 90 | 
 91 | class UpConv(Decoder):
 92 |     # UpConv decoder consists of 4 upconv modules with decreasing number of channels and increasing feature map size
 93 |     def upconv_module(self, in_channels):
 94 |         # UpConv module: unpool -> 5*5 conv -> batchnorm -> ReLU
 95 |         upconv = nn.Sequential(collections.OrderedDict([
 96 |             ('unpool', Unpool(in_channels)),
 97 |             ('conv', nn.Conv2d(in_channels, in_channels // 2, kernel_size=5, stride=1, padding=2, bias=False)),
 98 |             ('batchnorm', nn.BatchNorm2d(in_channels // 2)),
 99 |             ('relu', nn.ReLU()),
100 |         ]))
101 |         return upconv
102 | 
103 |     def __init__(self, in_channels):
104 |         super(UpConv, self).__init__()
105 |         self.layer1 = self.upconv_module(in_channels)
106 |         self.layer2 = self.upconv_module(in_channels // 2)
107 |         self.layer3 = self.upconv_module(in_channels // 4)
108 |         self.layer4 = self.upconv_module(in_channels // 8)
109 | 
110 | 
111 | class FasterUpConv(Decoder):
112 |     # Faster Upconv using pixelshuffle
113 | 
114 |     class faster_upconv_module(nn.Module):
115 | 
116 |         def __init__(self, in_channel):
117 |             super(FasterUpConv.faster_upconv_module, self).__init__()
118 | 
119 |             self.conv1_ = nn.Sequential(collections.OrderedDict([
120 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=3)),
121 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
122 |             ]))
123 | 
124 |             self.conv2_ = nn.Sequential(collections.OrderedDict([
125 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(2, 3))),
126 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
127 |             ]))
128 | 
129 |             self.conv3_ = nn.Sequential(collections.OrderedDict([
130 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(3, 2))),
131 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
132 |             ]))
133 | 
134 |             self.conv4_ = nn.Sequential(collections.OrderedDict([
135 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=2)),
136 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
137 |             ]))
138 | 
139 |             self.ps = nn.PixelShuffle(2)
140 |             self.relu = nn.ReLU(inplace=True)
141 | 
142 |         def forward(self, x):
143 |             # print('Upmodule x size = ', x.size())
144 |             x1 = self.conv1_(nn.functional.pad(x, (1, 1, 1, 1)))
145 |             x2 = self.conv2_(nn.functional.pad(x, (1, 1, 0, 1)))
146 |             x3 = self.conv3_(nn.functional.pad(x, (0, 1, 1, 1)))
147 |             x4 = self.conv4_(nn.functional.pad(x, (0, 1, 0, 1)))
148 | 
149 |             x = torch.cat((x1, x2, x3, x4), dim=1)
150 | 
151 |             output = self.ps(x)
152 |             output = self.relu(output)
153 | 
154 |             return output
155 | 
156 |     def __init__(self, in_channel):
157 |         super(FasterUpConv, self).__init__()
158 | 
159 |         self.layer1 = self.faster_upconv_module(in_channel)
160 |         self.layer2 = self.faster_upconv_module(in_channel // 2)
161 |         self.layer3 = self.faster_upconv_module(in_channel // 4)
162 |         self.layer4 = self.faster_upconv_module(in_channel // 8)
163 | 
164 | 
165 | class UpProj(Decoder):
166 |     # UpProj decoder consists of 4 upproj modules with decreasing number of channels and increasing feature map size
167 | 
168 |     class UpProjModule(nn.Module):
169 |         # UpProj module has two branches, with a Unpool at the start and a ReLu at the end
170 |         #   upper branch: 5*5 conv -> batchnorm -> ReLU -> 3*3 conv -> batchnorm
171 |         #   bottom branch: 5*5 conv -> batchnorm
172 | 
173 |         def __init__(self, in_channels):
174 |             super(UpProj.UpProjModule, self).__init__()
175 |             out_channels = in_channels // 2
176 |             self.unpool = Unpool(in_channels)
177 |             self.upper_branch = nn.Sequential(collections.OrderedDict([
178 |                 ('conv1', nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)),
179 |                 ('batchnorm1', nn.BatchNorm2d(out_channels)),
180 |                 ('relu', nn.ReLU()),
181 |                 ('conv2', nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)),
182 |                 ('batchnorm2', nn.BatchNorm2d(out_channels)),
183 |             ]))
184 |             self.bottom_branch = nn.Sequential(collections.OrderedDict([
185 |                 ('conv', nn.Conv2d(in_channels, out_channels, kernel_size=5, stride=1, padding=2, bias=False)),
186 |                 ('batchnorm', nn.BatchNorm2d(out_channels)),
187 |             ]))
188 |             self.relu = nn.ReLU()
189 | 
190 |         def forward(self, x):
191 |             x = self.unpool(x)
192 |             x1 = self.upper_branch(x)
193 |             x2 = self.bottom_branch(x)
194 |             x = x1 + x2
195 |             x = self.relu(x)
196 |             return x
197 | 
198 |     def __init__(self, in_channels):
199 |         super(UpProj, self).__init__()
200 |         self.layer1 = self.UpProjModule(in_channels)
201 |         self.layer2 = self.UpProjModule(in_channels // 2)
202 |         self.layer3 = self.UpProjModule(in_channels // 4)
203 |         self.layer4 = self.UpProjModule(in_channels // 8)
204 | 
205 | 
206 | class FasterUpProj(Decoder):
207 |     # Faster UpProj decorder using pixelshuffle
208 | 
209 |     class faster_upconv(nn.Module):
210 | 
211 |         def __init__(self, in_channel):
212 |             super(FasterUpProj.faster_upconv, self).__init__()
213 | 
214 |             self.conv1_ = nn.Sequential(collections.OrderedDict([
215 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=3)),
216 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
217 |             ]))
218 | 
219 |             self.conv2_ = nn.Sequential(collections.OrderedDict([
220 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(2, 3))),
221 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
222 |             ]))
223 | 
224 |             self.conv3_ = nn.Sequential(collections.OrderedDict([
225 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=(3, 2))),
226 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
227 |             ]))
228 | 
229 |             self.conv4_ = nn.Sequential(collections.OrderedDict([
230 |                 ('conv1', nn.Conv2d(in_channel, in_channel // 2, kernel_size=2)),
231 |                 ('bn1', nn.BatchNorm2d(in_channel // 2)),
232 |             ]))
233 | 
234 |             self.ps = nn.PixelShuffle(2)
235 |             self.relu = nn.ReLU(inplace=True)
236 | 
237 |         def forward(self, x):
238 |             # print('Upmodule x size = ', x.size())
239 |             x1 = self.conv1_(nn.functional.pad(x, (1, 1, 1, 1)))
240 |             x2 = self.conv2_(nn.functional.pad(x, (1, 1, 0, 1)))
241 |             x3 = self.conv3_(nn.functional.pad(x, (0, 1, 1, 1)))
242 |             x4 = self.conv4_(nn.functional.pad(x, (0, 1, 0, 1)))
243 |             # print(x1.size(), x2.size(), x3.size(), x4.size())
244 | 
245 |             x = torch.cat((x1, x2, x3, x4), dim=1)
246 | 
247 |             x = self.ps(x)
248 |             return x
249 | 
250 |     class FasterUpProjModule(nn.Module):
251 |         def __init__(self, in_channels):
252 |             super(FasterUpProj.FasterUpProjModule, self).__init__()
253 |             out_channels = in_channels // 2
254 | 
255 |             self.upper_branch = nn.Sequential(collections.OrderedDict([
256 |                 ('faster_upconv', FasterUpProj.faster_upconv(in_channels)),
257 |                 ('relu', nn.ReLU(inplace=True)),
258 |                 ('conv', nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)),
259 |                 ('batchnorm', nn.BatchNorm2d(out_channels)),
260 |             ]))
261 |             self.bottom_branch = FasterUpProj.faster_upconv(in_channels)
262 |             self.relu = nn.ReLU(inplace=True)
263 | 
264 |         def forward(self, x):
265 |             x1 = self.upper_branch(x)
266 |             x2 = self.bottom_branch(x)
267 |             x = x1 + x2
268 |             x = self.relu(x)
269 |             return x
270 | 
271 |     def __init__(self, in_channel):
272 |         super(FasterUpProj, self).__init__()
273 | 
274 |         self.layer1 = self.FasterUpProjModule(in_channel)
275 |         self.layer2 = self.FasterUpProjModule(in_channel // 2)
276 |         self.layer3 = self.FasterUpProjModule(in_channel // 4)
277 |         self.layer4 = self.FasterUpProjModule(in_channel // 8)
278 | 
279 | 
280 | def choose_decoder(decoder, in_channels):
281 |     if decoder[:6] == 'deconv':
282 |         assert len(decoder) == 7
283 |         kernel_size = int(decoder[6])
284 |         return DeConv(in_channels, kernel_size)
285 |     elif decoder == "upproj":
286 |         return UpProj(in_channels)
287 |     elif decoder == "upconv":
288 |         return UpConv(in_channels)
289 |     elif decoder == "fasterupproj":
290 |         return FasterUpProj(in_channels)
291 |     else:
292 |         assert False, "invalid option for decoder: {}".format(decoder)
293 | 
294 | class ResNet(nn.Module):
295 |     def __init__(self, dataset = 'kitti', layers = 50, decoder = 'upproj', output_size=(228, 304), in_channels=3, pretrained=True):
296 | 
297 |         if layers not in [18, 34, 50, 101, 152]:
298 |             raise RuntimeError(
299 |                 'Only 18, 34, 50, 101, and 152 layer model are defined for ResNet. Got {}'.format(layers))
300 | 
301 |         super(ResNet, self).__init__()
302 |         pretrained_model = torchvision.models.__dict__['resnet{}'.format(layers)](pretrained=pretrained)
303 | 
304 |         if in_channels == 3:
305 |             self.conv1 = pretrained_model._modules['conv1']
306 |             self.bn1 = pretrained_model._modules['bn1']
307 |         else:
308 |             self.conv1 = nn.Conv2d(in_channels, 64, kernel_size=7, stride=2, padding=3, bias=False)
309 |             self.bn1 = nn.BatchNorm2d(64)
310 |             weights_init(self.conv1)
311 |             weights_init(self.bn1)
312 | 
313 |         self.output_size = output_size
314 | 
315 |         self.relu = pretrained_model._modules['relu']
316 |         self.maxpool = pretrained_model._modules['maxpool']
317 |         self.layer1 = pretrained_model._modules['layer1']
318 |         self.layer2 = pretrained_model._modules['layer2']
319 |         self.layer3 = pretrained_model._modules['layer3']
320 |         self.layer4 = pretrained_model._modules['layer4']
321 | 
322 |         # clear memory
323 |         del pretrained_model
324 | 
325 |         # define number of intermediate channels
326 |         if layers <= 34:
327 |             num_channels = 512
328 |         elif layers >= 50:
329 |             num_channels = 2048
330 | 
331 |         self.conv2 = nn.Conv2d(num_channels, num_channels // 2, kernel_size=1, bias=False)
332 |         self.bn2 = nn.BatchNorm2d(num_channels // 2)
333 | 
334 |         self.upSample = choose_decoder(decoder, num_channels // 2)
335 | 
336 |         # setting bias=true doesn't improve accuracy
337 |         self.conv3 = nn.Conv2d(num_channels // 32, 3, kernel_size=3, stride=1, padding=1, bias=False)
338 |         self.bilinear = nn.Upsample(size=self.output_size, mode='bilinear', align_corners=True)
339 | 
340 |         # weight init
341 |         self.conv2.apply(weights_init)
342 |         self.bn2.apply(weights_init)
343 | 
344 |         self.upSample.apply(weights_init)
345 | 
346 |         self.conv3.apply(weights_init)
347 | 
348 |     def forward(self, x):
349 |         # resnet
350 |         x = self.conv1(x)
351 |         x = self.bn1(x)
352 |         x = self.relu(x)
353 |         x = self.maxpool(x)
354 |         x1 = self.layer1(x)
355 |         x2 = self.layer2(x1)
356 |         x3 = self.layer3(x2)
357 |         x4 = self.layer4(x3)
358 | 
359 |         x = self.conv2(x4)
360 |         x = self.bn2(x)
361 | 
362 |         x = self.upSample(x)
363 | 
364 |         x = self.conv3(x)
365 |         x = self.bilinear(x)
366 | 
367 |         return x
368 | 
369 |     def get_1x_lr_params(self):
370 |         """
371 |         This generator returns all the parameters of the net layer whose learning rate is 1x lr.
372 |         """
373 |         b = [self.conv1, self.bn1, self.relu, self.maxpool, self.layer1, self.layer2, self.layer3, self.layer4]
374 |         for i in range(len(b)):
375 |             for k in b[i].parameters():
376 |                 if k.requires_grad:
377 |                     yield k
378 | 
379 |     def get_10x_lr_params(self):
380 |         """
381 |         This generator returns all the parameters of the net layer whose learning rate is 20x lr.
382 |         """
383 |         b = [self.conv2, self.bn2, self.upSample, self.conv3, self.bilinear]
384 |         for j in range(len(b)):
385 |             for k in b[j].parameters():
386 |                 if k.requires_grad:
387 |                     yield k
388 | 
389 | 
390 | class Discriminator(nn.Module):
391 |     """
392 |     Discriminator for FCRN
393 |     """
394 |     def __init__(self, n_in):
395 |         super(Discriminator, self).__init__()
396 |         self.n_in = n_in
397 |         self.modules = []
398 | 
399 |         self.modules += [   nn.Conv2d(self.n_in, 32, kernel_size=3, stride=(1, 1), padding=(2, 2)),
400 |                             nn.MaxPool2d(kernel_size=(2, 2)),
401 |                             nn.LeakyReLU(0.2, True),
402 |                             nn.Conv2d(32, 32, 3, 1, 1),
403 |                             nn.MaxPool2d(kernel_size=(2,2)),
404 |                             nn.LeakyReLU(0.2, True),
405 |                             nn.Conv2d(32, 32, 3, 1, 1),
406 |                             nn.MaxPool2d(kernel_size=(2,2)),
407 |                             nn.LeakyReLU(0.2, True),
408 |                             nn.Conv2d(32, 32, 3, 1, 1),
409 |                             nn.MaxPool2d(kernel_size=(2,2)),
410 |                             nn.LeakyReLU(0.2, True),
411 |         ]
412 | 
413 |         self.model = nn.Sequential(*self.modules)
414 |         self.fc = nn.Linear(32 * 4, 1)
415 |         
416 |     def forward(self, x):
417 |         x = self.model(x)
418 |         x = x.view(x.size(0), -1)
419 |         x = self.fc(x)
420 |         return torch.sigmoid(x)
421 | 
422 | 
423 | """
424 | Define Loss functions here
425 | """
426 | class ScaleInvariantError(nn.Module):
427 |     """
428 |     Scale invariant error defined in Eigen's paper!
429 |     """
430 | 
431 |     def __init__(self, lamada=0.5):
432 |         super(ScaleInvariantError, self).__init__()
433 |         self.lamada = lamada
434 |         return
435 | 
436 |     def forward(self, y_true, y_pred):
437 |         first_log = torch.log(torch.clamp(y_pred, 0, 1))
438 |         second_log = torch.log(torch.clamp(y_true, 0, 1))
439 |         d = first_log - second_log
440 |         loss = torch.mean(d * d) - self.lamada * torch.mean(d) * torch.mean(d)
441 |         return loss
442 | 
443 | 
444 | class MaskedMSELoss(nn.Module):
445 |     def __init__(self):
446 |         super(MaskedMSELoss, self).__init__()
447 | 
448 |     def forward(self, pred, target):
449 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
450 |         valid_mask = (target > 0).detach()
451 |         diff = target - pred
452 |         diff = diff[valid_mask]
453 |         self.loss = (diff ** 2).mean()
454 |         return self.loss
455 | 
456 | 
457 | class MaskedL1Loss(nn.Module):
458 |     def __init__(self):
459 |         super(MaskedL1Loss, self).__init__()
460 | 
461 |     def forward(self, pred, target):
462 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
463 |         valid_mask = (target > 0).detach()
464 |         diff = target - pred
465 |         diff = diff[valid_mask]
466 |         self.loss = diff.abs().mean()
467 |         return self.loss
468 | 
469 | 
470 | class berHuLoss(nn.Module):
471 |     def __init__(self):
472 |         super(berHuLoss, self).__init__()
473 | 
474 |     def forward(self, pred, target):
475 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
476 | 
477 |         huber_c = torch.max(pred - target)
478 |         huber_c = 0.2 * huber_c
479 | 
480 |         valid_mask = (target > 0).detach()
481 |         diff = target - pred
482 |         diff = diff[valid_mask]
483 |         diff = diff.abs()
484 | 
485 |         huber_mask = (diff > huber_c).detach()
486 | 
487 |         diff2 = diff[huber_mask]
488 |         diff2 = diff2 ** 2
489 | 
490 |         self.loss = torch.cat((diff, diff2)).mean()
491 | 
492 |         return self.loss
493 | 
494 |     
495 | def set_requires_grad(net, requires_grad=False):
496 |     if net is not None:
497 |         for param in net.parameters():
498 |             param.requires_grad = requires_grad
499 | 
500 | def save_image(model, x, batch, mode="train"):
501 |     pred = model(x)
502 |     pred *= 100
503 |     npimg = pred.cpu().detach().numpy()
504 |     npimg = np.transpose(npimg, (0, 2, 3, 1))
505 |     cv2.imwrite('saved_images/%s_image_%d.jpg'%(mode, batch), npimg[0])
506 | 
507 | 
508 | def train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, 
509 |           criterion_berHu, criterion_GAN, optimizer, optimizer_D, epoch, batch_size, val_frequency=10):
510 |     model.train()  # switch to train mode
511 |     eval_mode = False
512 |     num_batches = len(train_loader)// batch_size
513 |     init_lr = optimizer.param_groups[0]['lr']
514 |     
515 |     total_losses_log = {"l1":0, "mse":0, "berHu":0, "adv":0}
516 |     
517 |     valid_T = torch.ones(batch_size, 1).cuda()
518 |     zeros_T = torch.zeros(batch_size, 1).cuda()
519 |     
520 |     train_loss_list = []
521 |     val_loss_list = []
522 |     
523 |     for iter_ in range(num_batches):
524 |         # Adjust Learning Rate
525 |         #if iter_ % 1000 == 0 and optimizer.param_groups[0]['lr'] > init_lr/10.0:
526 |         #    for param_group in optimizer.param_groups:
527 |         #        param_group['lr'] *= 0.98
528 |         
529 |         input, target = next(train_loader.get_one_batch(batch_size))
530 |         input, target = input.float(), target.float()
531 |         input, target = input.cuda(), target.cuda()
532 |         torch.cuda.synchronize()
533 |         
534 |         optimizer.zero_grad()
535 |         
536 |         pred = model(input)
537 |         
538 |         loss_L1 = criterion_L1(pred, target)
539 |         loss_MSE = criterion_MSE(pred, target)
540 |         loss_berHu = criterion_berHu(pred, target)
541 |         #loss_SI = criterion_SI(pred, target)
542 |         
543 |         set_requires_grad(discriminator, False)
544 |         
545 |         loss_adv = 0
546 |         
547 |         for a in range(3):
548 |             for b in range(9):
549 |                 row = 30 * a
550 |                 col = 30 * b
551 |                 patch_fake = pred[:, :, row:row+30, col:col+30]
552 |                 pred_fake = discriminator(patch_fake)
553 |                 loss_adv += criterion_GAN(pred_fake, valid_T)
554 |     
555 |         loss_gen = loss_L1 + loss_adv
556 |         loss_gen.backward()
557 |         optimizer.step()
558 |         
559 |         set_requires_grad(discriminator, True)
560 |         optimizer_D.zero_grad()
561 |         loss_D = 0
562 |         for a in range(3):
563 |             for b in range(9):
564 |                 row = 30 * a
565 |                 col = 30 * b
566 |                 patch_fake = pred[:, :, row:row+30, col:col+30]
567 |                 patch_real = target[:, :, row:row+30, col:col+30]
568 |                 pred_fake = discriminator(patch_fake.detach())
569 |                 pred_real = discriminator(patch_real)
570 |                 loss_D_fake = criterion_GAN(pred_fake, zeros_T)
571 |                 loss_D_real = criterion_GAN(pred_real, valid_T)
572 |                 loss_D += 0.5 * (loss_D_fake + loss_D_real)
573 |         
574 |         loss_D.backward()
575 |         optimizer_D.step()
576 | 
577 |         torch.cuda.synchronize()
578 |         
579 |         total_losses_log["l1"] += loss_L1.item()
580 |         total_losses_log["mse"] += loss_MSE.item()
581 |         total_losses_log["berHu"] += loss_berHu.item()
582 |         total_losses_log["adv"] += loss_adv.item()
583 |          
584 |         if (iter_ + 1) % 10 == 0:
585 |             save_image(model, input, iter_)
586 |             print('Train Epoch: {} Batch: [{}/{}], ADV:{:0.3f} L1 ={:0.3f}, MSE={:0.3f}, berHu={:0.3f}, Disc:{:0.3f}'.format(
587 |                 epoch, iter_ + 1, num_batches, loss_adv.item(),
588 |                 loss_L1.item(), loss_MSE.item(), loss_berHu.item(), loss_D.item()))
589 |         
590 |         if (iter_ + 1) % val_frequency == 0:
591 |             val_loss = validate(val_loader, model, discriminator, criterion_L1, criterion_MSE, criterion_berHu, criterion_GAN, batch_size, iter_)
592 |             for key in total_losses_log.keys():
593 |                 total_losses_log[key] /= (val_frequency * batch_size)
594 |             train_loss_list.append(total_losses_log)
595 |             val_loss_list.append(val_loss)
596 |             save_losses(train_loss_list, val_loss_list)
597 |             print("TRAIN:- ", total_losses_log)
598 |             print("VAL :- ", val_loss)
599 |             total_losses_log = {"l1":0, "mse":0, "berHu":0, "adv":0}
600 | 
601 | def save_losses(train_loss_list, val_loss_list):
602 |     out = {"train_loss": train_loss_list, "val_loss":val_loss_list}
603 |     with open("losses.pkl", "wb") as log:
604 |         pickle.dump(out, log)
605 |         
606 | def validate(val_loader, model, discriminator, criterion_L1, criterion_MSE, 
607 |              criterion_berHu, criterion_GAN, batch_size, train_iter):
608 |     model.eval()  # switch to evaluate mode
609 |     discriminator.eval()
610 |     
611 |     valid_T = torch.ones(batch_size, 1).cuda()
612 |     zeros_T = torch.zeros(batch_size, 1).cuda()
613 |     
614 |     #num_batches = len(val_loader) // batch_size
615 |     total_losses = {"l1":0, "mse":0, "berHu":0, "adv":0}
616 |     num_batches = 25
617 |     for i in range(num_batches):
618 | 
619 |         input, target = next(val_loader.get_one_batch(batch_size))
620 |         input, target = input.float(), target.float()
621 |         input, target = input.cuda(), target.cuda()
622 |         
623 |         torch.cuda.synchronize()
624 | 
625 |         # compute output
626 |         end = time.time()
627 |         with torch.no_grad():
628 |             pred = model(input)
629 |             
630 |         torch.cuda.synchronize()
631 |         
632 |         loss_L1 = criterion_L1(pred, target)
633 |         loss_MSE = criterion_MSE(pred, target)
634 |         loss_berHu = criterion_berHu(pred, target)
635 |         
636 |         loss_adv = 0
637 |         with torch.no_grad():
638 |             for a in range(3):
639 |                 for b in range(9):
640 |                     row = 30 * a
641 |                     col = 30 * b
642 |                     patch_fake = pred[:, :, row:row+30, col:col+30]
643 |                     pred_fake = discriminator(patch_fake)
644 |                     loss_adv += criterion_GAN(pred_fake, valid_T)
645 |                     
646 |         total_losses["l1"] += loss_L1.item()
647 |         total_losses["mse"] += loss_MSE.item()
648 |         total_losses["berHu"] += loss_berHu.item()
649 |         total_losses["adv"] += loss_adv.item()
650 |         
651 |         if i==0:
652 |             save_image(model, input, train_iter, "val")
653 |     for key in total_losses.keys():
654 |         total_losses[key] /= (num_batches * batch_size)
655 |     model.train()
656 |     discriminator.train() 
657 |     return total_losses
658 | 
659 | raw_data_dir = ""
660 | depth_maps_dir = ""
661 | print("=> Loading Data ...")
662 | train_loader = DataLoader(raw_data_dir, depth_maps_dir)
663 | val_loader = DataLoader(raw_data_dir, depth_maps_dir, mode = "val")
664 | 
665 | print("=> creating Model")
666 | model = ResNet(layers=152, output_size=(90, 270), pretrained=True)
667 | discriminator = Discriminator(3)
668 | discriminator.apply(weights_init)
669 | 
670 | print("=> model created.")
671 | start_epoch = 0
672 | init_lr = 0.001
673 | 
674 | train_params = [{'params': model.get_1x_lr_params(), 'lr': init_lr},
675 |                 {'params': model.get_10x_lr_params(), 'lr': init_lr * 10}]
676 | 
677 | optimizer = torch.optim.SGD(train_params, lr=init_lr, weight_decay=4e-5)
678 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=init_lr)
679 | 
680 | # You can use DataParallel() whether you use Multi-GPUs or not
681 | model = nn.DataParallel(model).cuda()
682 | discriminator = discriminator.cuda()
683 | 
684 | # Define Loss Function
685 | criterion_L1 = MaskedL1Loss()
686 | criterion_berHu = berHuLoss()
687 | # criterion_SI = ScaleInvariantError()
688 | criterion_MSE = MaskedMSELoss()
689 | criterion_GAN = nn.BCELoss()
690 | 
691 | batch_size = 64
692 | val_frequency = 100
693 | 
694 | for epoch in range(10):
695 |     # Train the Model
696 |     train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, criterion_berHu, 
697 |           criterion_GAN, optimizer, optimizer_D, epoch, batch_size, val_frequency)
698 |      
699 |     # Save Checkpoint
700 |     torch.save({
701 |             'epoch': epoch,
702 |             'model_state_dict': model.state_dict(),
703 |             'optimizer_state_dict': optimizer.state_dict(),
704 |             }, './ResNet.pth')
705 | 


--------------------------------------------------------------------------------
/train_dorn.py:
--------------------------------------------------------------------------------
  1 | from datetime import datetime
  2 | import shutil
  3 | import socket
  4 | import time
  5 | import torch
  6 | from tensorboardX import SummaryWriter
  7 | from torch.optim import lr_scheduler
  8 | 
  9 | from data_loader import KittiData
 10 | from dorn.evaluation import AverageMeter, Result
 11 | import utils
 12 | import dorn.criterion as criteria
 13 | import os
 14 | import torch.nn as nn
 15 | 
 16 | from dorn.networks import DORN
 17 | 
 18 | 
 19 | # set device
 20 | device = "cuda" if torch.cuda.is_available() else "cpu"
 21 | 
 22 | best_result = Result()
 23 | best_result.set_to_worst()
 24 | 
 25 | raw_data_dir = "data/"
 26 | depth_maps_dir = "data/depth_maps/"
 27 | output_directory = "checkpoints/"
 28 | epoch_file = "epoch.txt"
 29 | 
 30 | 
 31 | def create_loader(opts):
 32 |     train_set = KittiData(raw_data_dir, depth_maps_dir)
 33 |     val_set = KittiData(raw_data_dir, depth_maps_dir, mode='val')
 34 |     train_loader = torch.utils.data.DataLoader(train_set, batch_size=opts.batch_size, shuffle=True, pin_memory=True)
 35 |     val_loader = torch.utils.data.DataLoader(val_set, batch_size=1, shuffle=False, pin_memory=True)
 36 |     return train_loader, val_loader
 37 | 
 38 | 
 39 | def main():
 40 |     global best_result
 41 |     model = DORN()
 42 |     opts = utils.get_opts()
 43 |     epoch_tracker = utils.EpochTracker(epoch_file)
 44 | 
 45 |     train_loader, val_loader = create_loader(opts)
 46 | 
 47 |     name = output_directory + 'checkpoint-' + str(epoch_tracker.epoch) + '.pth.tar'
 48 |     if os.path.exists(name):
 49 |         checkpoint = torch.load(name)
 50 | 
 51 |         start_epoch = checkpoint['epoch'] + 1
 52 |         best_result = checkpoint['best_result']
 53 |         optimizer = checkpoint['optimizer']
 54 |         iteration = checkpoint['iteration']
 55 | 
 56 |         # solve 'out of memory'
 57 |         model = checkpoint['model']
 58 | 
 59 |         print("=> loaded checkpoint (epoch {})".format(checkpoint['epoch']))
 60 | 
 61 |         # clear memory
 62 |         del checkpoint
 63 |         # del model_dict
 64 |         if torch.cuda.is_available():
 65 |             torch.cuda.empty_cache()
 66 |     else:
 67 |         start_epoch = 0
 68 |         iteration = 0
 69 |         # different modules have different learning rate
 70 |         train_params = [{'params': model.get_1x_lr_params(), 'lr': opts.lr},
 71 |                         {'params': model.get_10x_lr_params(), 'lr': opts.lr * 10}]
 72 | 
 73 |         optimizer = torch.optim.SGD(train_params, lr=opts.lr, momentum=opts.momentum, weight_decay=opts.weight_decay)
 74 | 
 75 |         # You can use DataParallel() whether you use Multi-GPUs or not
 76 |         if torch.cuda.is_available():
 77 |             model = nn.DataParallel(model).cuda()
 78 | 
 79 |     # when training, use reduceLROnPlateau to reduce learning rate
 80 |     scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=opts.lr_patience)
 81 | 
 82 |     # loss function
 83 |     criterion = criteria.ordLoss(device)
 84 | 
 85 |     if not os.path.exists(output_directory):
 86 |         os.makedirs(output_directory)
 87 | 
 88 |     best_txt = os.path.join(output_directory, 'best.txt')
 89 |     config_txt = os.path.join(output_directory, 'config.txt')
 90 | 
 91 |     # create log
 92 |     log_path = os.path.join(output_directory, 'logs',
 93 |                             datetime.now().strftime('%b%d_%H-%M-%S') + '_' + socket.gethostname())
 94 |     if os.path.isdir(log_path):
 95 |         shutil.rmtree(log_path)
 96 |     os.makedirs(log_path)
 97 |     logger = SummaryWriter(log_path)
 98 | 
 99 |     for epoch in range(start_epoch, opts.epochs):
100 | 
101 |         # remember change of the learning rate
102 |         for i, param_group in enumerate(optimizer.param_groups):
103 |             old_lr = float(param_group['lr'])
104 |             logger.add_scalar('Lr/lr_' + str(i), old_lr, epoch)
105 | 
106 |         train(train_loader, model, criterion, optimizer, epoch, logger, device, opts, iteration)  # train for one epoch
107 |         result, img_merge = validate(val_loader, model, epoch, logger, opts)  # evaluate on validation set
108 |         iteration = 0
109 |         # remember best rmse and save checkpoint
110 |         is_best = result.rmse < best_result.rmse
111 |         if is_best:
112 |             best_result = result
113 |             with open(best_txt, 'w') as txtfile:
114 |                 txtfile.write(
115 |                     "epoch={}, rmse={:.3f}, rml={:.3f}, log10={:.3f}, d1={:.3f}, d2={:.3f}, dd31={:.3f}, "
116 |                     "t_gpu={:.4f}".
117 |                         format(epoch, result.rmse, result.absrel, result.lg10, result.delta1, result.delta2,
118 |                                result.delta3,
119 |                                result.gpu_time))
120 |             if img_merge is not None:
121 |                 img_filename = output_directory + '/comparison_best.png'
122 |                 utils.save_image(img_merge, img_filename)
123 | 
124 |         # save checkpoint for each epoch
125 |         utils.save_checkpoint({
126 |             'args': opts,
127 |             'epoch': epoch,
128 |             'model': model,
129 |             'best_result': best_result,
130 |             'optimizer': optimizer,
131 |             'iteration': iteration
132 |         }, is_best, epoch, output_directory)
133 | 
134 |         epoch_tracker.write(epoch)
135 |         # when rml doesn't fall, reduce learning rate
136 |         scheduler.step(result.absrel)
137 | 
138 |     logger.close()
139 | 
140 | 
141 | # train
142 | def train(train_loader, model, criterion, optimizer, epoch, logger, device, opts, iteration):
143 |     average_meter = AverageMeter()
144 |     model.train()  # switch to train mode
145 |     end = time.time()
146 | 
147 |     batch_num = len(train_loader)
148 | 
149 |     for i, (input, target) in enumerate(train_loader):
150 |         if i < iteration:
151 |             continue
152 | 
153 |         input = input.to(device)
154 |         target = target.to(device)
155 | 
156 |         if torch.cuda.is_available():
157 |             torch.cuda.synchronize()
158 |         data_time = time.time() - end
159 | 
160 |         # compute pred
161 |         end = time.time()
162 | 
163 |         with torch.autograd.detect_anomaly():
164 |             pred_d, pred_ord = model(input) 
165 |             target_c = utils.get_labels_sid(opts, target, device)  # using sid, discretize the groundtruth
166 |             loss = criterion(pred_ord, target_c)
167 |             optimizer.zero_grad()
168 |             loss.backward()  # compute gradient and do SGD step
169 |             optimizer.step()
170 |         
171 |         if torch.cuda.is_available():
172 |             torch.cuda.synchronize()
173 |         
174 |         gpu_time = time.time() - end
175 | 
176 |         # measure accuracy and record loss
177 |         result = Result()
178 |         depth = utils.get_depth_sid(opts, pred_d, device)
179 |         result.evaluate(depth.data, target.data)
180 |         average_meter.update(result, gpu_time, data_time, input.size(0))
181 |         end = time.time()
182 | 
183 |         if (i + 1) % opts.save_freq == 0:
184 |             utils.save_checkpoint({
185 |             'args': opts,
186 |             'epoch': epoch,
187 |             'model': model,
188 |             'best_result': best_result,
189 |             'optimizer': optimizer,
190 |             'iteration':i+1
191 |             }, False, epoch, output_directory)
192 | 
193 |         if (i + 1) % opts.print_freq == 0:
194 |             print('Train Epoch: {0} [{1}/{2}]\t'
195 |                   't_Data={data_time:.3f}({average.data_time:.3f}) '
196 |                   't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
197 |                   'Loss={Loss:.5f} '
198 |                   'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
199 |                   'RML={result.absrel:.2f}({average.absrel:.2f}) '
200 |                   'Log10={result.lg10:.3f}({average.lg10:.3f}) '
201 |                   'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
202 |                   'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
203 |                   'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
204 |                 epoch, i + 1, len(train_loader), data_time=data_time,
205 |                 gpu_time=gpu_time, Loss=loss.item(), result=result, average=average_meter.average()))
206 |             current_step = epoch * batch_num + i
207 |             logger.add_scalar('Train/RMSE', result.rmse, current_step)
208 |             logger.add_scalar('Train/rml', result.absrel, current_step)
209 |             logger.add_scalar('Train/Log10', result.lg10, current_step)
210 |             logger.add_scalar('Train/Delta1', result.delta1, current_step)
211 |             logger.add_scalar('Train/Delta2', result.delta2, current_step)
212 |             logger.add_scalar('Train/Delta3', result.delta3, current_step)
213 | 
214 |         print("GPU:", torch.cuda.memory_allocated())
215 | 
216 |         if torch.cuda.is_available():
217 |             torch.cuda.empty_cache()
218 | 
219 |     avg = average_meter.average()
220 | 
221 | 
222 | # validation
223 | def validate(val_loader, model, epoch, logger, opts):
224 |     average_meter = AverageMeter()
225 | 
226 |     model.eval()  # switch to evaluate mode
227 | 
228 |     end = time.time()
229 | 
230 |     skip = len(val_loader) // 9  # save images every skip iters
231 | 
232 |     for i, (input, target) in enumerate(val_loader):
233 | 
234 |         input, target = input.cuda(), target.cuda()
235 |         torch.cuda.synchronize()
236 |         data_time = time.time() - end
237 | 
238 |         # compute output
239 |         end = time.time()
240 |         with torch.no_grad():
241 |             pred, _ = model(input)
242 | 
243 |         torch.cuda.synchronize()
244 |         gpu_time = time.time() - end
245 | 
246 |         # measure accuracy and record loss
247 |         result = Result()
248 |         pred = utils.get_depth_sid(opts, pred)
249 |         result.evaluate(pred.data, target.data)
250 | 
251 |         average_meter.update(result, gpu_time, data_time, input.size(0))
252 |         end = time.time()
253 | 
254 |         # save 8 images for visualization
255 |         rgb = input
256 | 
257 |         if i == 0:
258 |             img_merge = utils.merge_into_row(rgb, target, pred)
259 |         elif (i < 8 * skip) and (i % skip == 0):
260 |             row = utils.merge_into_row(rgb, target, pred)
261 |             img_merge = utils.add_row(img_merge, row)
262 |         elif i == 8 * skip:
263 |             filename = output_directory + '/comparison_' + str(epoch) + '.png'
264 |             utils.save_image(img_merge, filename)
265 | 
266 |         if (i + 1) % opts.print_freq == 0:
267 |             print('Test: [{0}/{1}]\t'
268 |                   't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\t'
269 |                   'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
270 |                   'RML={result.absrel:.2f}({average.absrel:.2f}) '
271 |                   'Log10={result.lg10:.3f}({average.lg10:.3f}) '
272 |                   'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
273 |                   'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
274 |                   'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
275 |                 i + 1, len(val_loader), gpu_time=gpu_time, result=result, average=average_meter.average()))
276 | 
277 |     avg = average_meter.average()
278 | 
279 |     print('\n*\n'
280 |           'RMSE={average.rmse:.3f}\n'
281 |           'Rel={average.absrel:.3f}\n'
282 |           'Log10={average.lg10:.3f}\n'
283 |           'Delta1={average.delta1:.3f}\n'
284 |           'Delta2={average.delta2:.3f}\n'
285 |           'Delta3={average.delta3:.3f}\n'
286 |           't_GPU={time:.3f}\n'.format(
287 |         average=avg, time=avg.gpu_time))
288 | 
289 |     logger.add_scalar('Test/rmse', avg.rmse, epoch)
290 |     logger.add_scalar('Test/Rel', avg.absrel, epoch)
291 |     logger.add_scalar('Test/log10', avg.lg10, epoch)
292 |     logger.add_scalar('Test/Delta1', avg.delta1, epoch)
293 |     logger.add_scalar('Test/Delta2', avg.delta2, epoch)
294 |     logger.add_scalar('Test/Delta3', avg.delta3, epoch)
295 |     return avg, img_merge
296 | 
297 | 
298 | if __name__ == '__main__':
299 |     main()
300 | 


--------------------------------------------------------------------------------
/train_patchgan.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import torch
  3 | import torch.nn as nn
  4 | import torch.nn.functional as F
  5 | import torchvision.models
  6 | import collections
  7 | import math
  8 | import time
  9 | import pickle
 10 | import matplotlib.pyplot as plt
 11 | from PIL import Image
 12 | from tensorboardX import SummaryWriter
 13 | cmap = plt.cm.jet
 14 | 
 15 | from nyu_dataloader import DataLoader
 16 | import model
 17 | from unet import UNet
 18 | 
 19 | logger = SummaryWriter("runs/run1")
 20 | 
 21 | def weights_init(m):
 22 |     # Initialize filters with Gaussian random weights
 23 |     if isinstance(m, nn.Conv2d):
 24 |         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
 25 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 26 |         if m.bias is not None:
 27 |             m.bias.data.zero_()
 28 |     elif isinstance(m, nn.ConvTranspose2d):
 29 |         n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
 30 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 31 |         if m.bias is not None:
 32 |             m.bias.data.zero_()
 33 |     elif isinstance(m, nn.BatchNorm2d):
 34 |         m.weight.data.fill_(1)
 35 |         m.bias.data.zero_()
 36 | 
 37 | class MaskedMSELoss(nn.Module):
 38 |     def __init__(self):
 39 |         super(MaskedMSELoss, self).__init__()
 40 | 
 41 |     def forward(self, pred, target):
 42 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 43 |         valid_mask = (target > 0).detach()
 44 |         diff = target - pred
 45 |         diff = diff[valid_mask]
 46 |         self.loss = (diff ** 2).mean()
 47 |         return self.loss
 48 | 
 49 | 
 50 | class MaskedL1Loss(nn.Module):
 51 |     def __init__(self):
 52 |         super(MaskedL1Loss, self).__init__()
 53 | 
 54 |     def forward(self, pred, target):
 55 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 56 |         valid_mask = (target > 0).detach()
 57 |         diff = target - pred
 58 |         diff = diff[valid_mask]
 59 |         self.loss = diff.abs().mean()
 60 |         return self.loss
 61 | 
 62 | 
 63 | class berHuLoss(nn.Module):
 64 |     def __init__(self):
 65 |         super(berHuLoss, self).__init__()
 66 | 
 67 |     def forward(self, pred, target):
 68 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 69 | 
 70 |         huber_c = torch.max(pred - target)
 71 |         huber_c = 0.2 * huber_c
 72 | 
 73 |         valid_mask = (target > 0).detach()
 74 |         diff = target - pred
 75 |         diff = diff[valid_mask]
 76 |         diff = diff.abs()
 77 | 
 78 |         huber_mask = (diff > huber_c).detach()
 79 | 
 80 |         diff2 = diff[huber_mask]
 81 |         diff2 = diff2 ** 2
 82 | 
 83 |         self.loss = torch.cat((diff, diff2)).mean()
 84 | 
 85 |         return self.loss
 86 | 
 87 | 
 88 | class ScaleInvariantError(nn.Module):
 89 |     def __init__(self, lamada=0.5):
 90 |         super(ScaleInvariantError, self).__init__()
 91 |         self.lamada = lamada
 92 |         return
 93 | 
 94 |     def forward(self, y_true, y_pred):
 95 |         first_log = torch.log(torch.clamp(y_pred, 0.0001))
 96 |         second_log = torch.log(torch.clamp(y_true, 0.0001))
 97 |         d = first_log - second_log
 98 |         loss = torch.mean(d * d) - self.lamada * torch.mean(d) * torch.mean(d)
 99 |         return loss
100 | 
101 | def create_depth_color(depth):
102 |     d_min = np.min(depth)
103 |     d_max = np.max(depth)
104 |     depth_relative = (depth - d_min) / (d_max - d_min)
105 |     depth = (255 * cmap(depth_relative)[:, :, :3])
106 |     return depth
107 | 
108 | 
109 | def save_image(model, x, y, batch, mode="train"):
110 |     pred = model(x)
111 |     npimg = pred.cpu().detach().numpy()
112 |     depth = create_depth_color(np.transpose(npimg[0], [1,2,0])[:, :, 0])
113 |     target = create_depth_color(np.transpose(y[0].cpu().numpy(), [1,2,0])[:, :, 0])
114 |     orig = 255 * np.transpose(x[0].cpu().numpy(), [1,2,0])
115 | 
116 |     img = np.concatenate((orig, target, depth), axis =1)
117 |     
118 |     img = Image.fromarray(img.astype('uint8'))
119 |     img.save('saved_images/%s_image_%d.jpg'%(mode, batch))
120 |     
121 | def adjust_learning_rate(optimizer, epoch, lr_init):
122 |     """Sets the learning rate to the initial LR decayed by 2 every 5 epochs"""
123 |     lr = lr_init * (0.5 ** (epoch // 5))
124 |     for param_group in optimizer.param_groups:
125 |         param_group['lr'] = lr
126 | 
127 | class Discriminator(nn.Module):
128 |     def __init__(self):
129 |         super(Discriminator, self).__init__()
130 |         self.model = nn.Sequential (
131 |                     nn.Conv2d(1, 32, 4, 1, 0),
132 |                     nn.BatchNorm2d(32),
133 |                     nn.ReLU(True),
134 |                     nn.MaxPool2d(2),
135 |                     nn.Conv2d(32, 32, 3, 2, 1),
136 |                     nn.BatchNorm2d(32),
137 |                     nn.ReLU(True),
138 |                     nn.MaxPool2d(2),
139 |                     nn.Conv2d(32, 32, 3, 2, 1),
140 |                     nn.BatchNorm2d(32),
141 |                     nn.ReLU(True)
142 |                 )
143 |         self.out = nn.Linear(32, 1)
144 |     
145 |     def forward(self, x):
146 |         x = self.model(x)
147 |         x = x.view(-1, 32)
148 |         x = self.out(x)
149 |         return torch.sigmoid(x)
150 | 
151 | 
152 | def set_requires_grad(net, requires_grad=False):
153 |     if net is not None:
154 |         for param in net.parameters():
155 |             param.requires_grad = requires_grad
156 | 
157 | 
158 | def train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, 
159 |           criterion_berHu, criterion_GAN, optimizer, optimizer_D, epoch, batch_size):
160 |     
161 |     model.train()  # switch to train mode
162 |     eval_mode = False
163 |     init_lr = optimizer.param_groups[0]['lr']
164 |     
165 |     valid_T = torch.ones(batch_size, 1).cuda().double()
166 |     zeros_T = torch.zeros(batch_size, 1).cuda().double()
167 |     
168 |     for iter_ in range(num_batches):
169 |         input, target = next(train_loader.get_one_batch(batch_size))
170 |         input, target = input, target
171 |         input, target = input.cuda(), target.cuda()
172 |         torch.cuda.synchronize()
173 |         
174 |         optimizer.zero_grad()
175 |         
176 |         pred = model(input)
177 |         
178 |         loss_L1 = criterion_L1(pred, target)
179 |         loss_MSE = criterion_MSE(pred, target)
180 |         loss_berHu = criterion_berHu(pred, target)
181 |         loss_SI = criterion_SI(pred, target)
182 | 
183 |         set_requires_grad(discriminator, False)
184 |         
185 |         loss_adv = 0
186 |         
187 |         for a in range(12):
188 |             for b in range(16):
189 |                 row = 19 * a
190 |                 col = 19 * b
191 |                 patch_fake = pred[:, :, row:row+19, col:col+19]
192 |                 pred_fake = discriminator(patch_fake)
193 |                 loss_adv += criterion_GAN(pred_fake, valid_T)
194 |     
195 |         loss_gen = loss_SI + 0.5 * loss_adv
196 |         loss_gen.backward()
197 |         optimizer.step()
198 |         
199 |         set_requires_grad(discriminator, True)
200 |         optimizer_D.zero_grad()
201 |         loss_D = 0
202 |         for a in range(12):
203 |             for b in range(16):
204 |                 row = 19 * a
205 |                 col = 19 * b
206 |                 patch_fake = pred[:, :, row:row+19, col:col+19]
207 |                 patch_real = target[:, :, row:row+19, col:col+19]
208 |                 pred_fake = discriminator(patch_fake.detach())
209 |                 pred_real = discriminator(patch_real)
210 |                 loss_D_fake = criterion_GAN(pred_fake, zeros_T)
211 |                 loss_D_real = criterion_GAN(pred_real, valid_T)
212 |                 loss_D += 0.5 * (loss_D_fake + loss_D_real)
213 |         
214 |         loss_D.backward()
215 |         optimizer_D.step()
216 | 
217 |         torch.cuda.synchronize()
218 |         if (iter_ + 1) % 10 == 0:
219 |             save_image(model, input, target, iter_)
220 |             logger.add_scalar('L1', loss_L1.item())
221 |             logger.add_scalar('MSE', loss_MSE.item())
222 |             logger.add_scalar('berHu', loss_berHu.item())
223 |             logger.add_scalar('SI', loss_SI.item())
224 |             
225 |             print('Train Epoch: {} Batch: [{}/{}], SI: {:0.4f}, ADV:{:0.3f} L1 ={:0.3f}, MSE={:0.3f}, berHu={:0.3f}, Disc:{:0.3f}'.format(
226 |                 epoch, iter_ + 1, num_batches, loss_SI.item(), loss_adv.item(),
227 |                 loss_L1.item(), loss_MSE.item(), loss_berHu.item(), loss_D.item()))
228 | 
229 | 
230 | train_loader = DataLoader("../cnn_depth_tensorflow/data/nyu_datasets/")
231 | val_loader = DataLoader("../cnn_depth_tensorflow/data/nyu_datasets/", mode="val")
232 | 
233 | model = UNet(3, 1).double()
234 | model = nn.DataParallel(model).cuda()
235 | 
236 | discriminator = Discriminator().double()
237 | discriminator.apply(weights_init)
238 | discriminator = discriminator.cuda()
239 | discriminator = nn.DataParallel(discriminator)
240 | 
241 | optimizer = torch.optim.Adam(model.parameters(), lr = 0.001, weight_decay=1e-4)
242 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr = 4 * 0.001)
243 | 
244 | criterion_L1 = MaskedL1Loss()
245 | criterion_berHu = berHuLoss()
246 | criterion_MSE = MaskedMSELoss()
247 | criterion_SI = ScaleInvariantError()
248 | criterion_GAN = nn.BCELoss()
249 | criterion = nn.L1Loss()
250 | 
251 | batch_size = 8
252 | num_epochs = 40
253 | num_batches = len(train_loader)//batch_size
254 | 
255 | for epoch in range(25):
256 |     adjust_learning_rate(optimizer, epoch, 0.001)
257 |     train(train_loader, val_loader, model, discriminator, criterion_L1, criterion_MSE, criterion_berHu, 
258 |           criterion_GAN, optimizer, optimizer_D, epoch, batch_size)
259 |      
260 |     torch.save({
261 |             'epoch': epoch,
262 |             'model_state_dict': model.state_dict(),
263 |             'optimizer_state_dict': optimizer.state_dict(),
264 |             }, './unet.pth')
265 | 
266 | 
267 | 


--------------------------------------------------------------------------------
/train_unet.py:
--------------------------------------------------------------------------------
  1 | import numpy as np
  2 | import torch
  3 | import torch.nn as nn
  4 | import torch.nn.functional as F
  5 | import torchvision.models
  6 | import collections
  7 | import math
  8 | import time
  9 | import pickle
 10 | import matplotlib.pyplot as plt
 11 | from PIL import Image
 12 | from tensorboardX import SummaryWriter
 13 | cmap = plt.cm.jet
 14 | 
 15 | from nyu_dataloader import DataLoader
 16 | import model
 17 | from unet import UNet
 18 | 
 19 | writer = SummaryWriter("runs/run1")
 20 | 
 21 | def weights_init(m):
 22 |     # Initialize filters with Gaussian random weights
 23 |     if isinstance(m, nn.Conv2d):
 24 |         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
 25 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 26 |         if m.bias is not None:
 27 |             m.bias.data.zero_()
 28 |     elif isinstance(m, nn.ConvTranspose2d):
 29 |         n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
 30 |         m.weight.data.normal_(0, math.sqrt(2. / n))
 31 |         if m.bias is not None:
 32 |             m.bias.data.zero_()
 33 |     elif isinstance(m, nn.BatchNorm2d):
 34 |         m.weight.data.fill_(1)
 35 |         m.bias.data.zero_()
 36 | 
 37 | class MaskedMSELoss(nn.Module):
 38 |     def __init__(self):
 39 |         super(MaskedMSELoss, self).__init__()
 40 | 
 41 |     def forward(self, pred, target):
 42 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 43 |         valid_mask = (target > 0).detach()
 44 |         diff = target - pred
 45 |         diff = diff[valid_mask]
 46 |         self.loss = (diff ** 2).mean()
 47 |         return self.loss
 48 | 
 49 | 
 50 | class MaskedL1Loss(nn.Module):
 51 |     def __init__(self):
 52 |         super(MaskedL1Loss, self).__init__()
 53 | 
 54 |     def forward(self, pred, target):
 55 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 56 |         valid_mask = (target > 0).detach()
 57 |         diff = target - pred
 58 |         diff = diff[valid_mask]
 59 |         self.loss = diff.abs().mean()
 60 |         return self.loss
 61 | 
 62 | 
 63 | class berHuLoss(nn.Module):
 64 |     def __init__(self):
 65 |         super(berHuLoss, self).__init__()
 66 | 
 67 |     def forward(self, pred, target):
 68 |         assert pred.dim() == target.dim(), "inconsistent dimensions"
 69 | 
 70 |         huber_c = torch.max(pred - target)
 71 |         huber_c = 0.2 * huber_c
 72 | 
 73 |         valid_mask = (target > 0).detach()
 74 |         diff = target - pred
 75 |         diff = diff[valid_mask]
 76 |         diff = diff.abs()
 77 | 
 78 |         huber_mask = (diff > huber_c).detach()
 79 | 
 80 |         diff2 = diff[huber_mask]
 81 |         diff2 = diff2 ** 2
 82 | 
 83 |         self.loss = torch.cat((diff, diff2)).mean()
 84 | 
 85 |         return self.loss
 86 | 
 87 | 
 88 | class ScaleInvariantError(nn.Module):
 89 |     def __init__(self, lamada=0.5):
 90 |         super(ScaleInvariantError, self).__init__()
 91 |         self.lamada = lamada
 92 |         return
 93 | 
 94 |     def forward(self, y_true, y_pred):
 95 |         first_log = torch.log(torch.clamp(y_pred, 0.0001))
 96 |         second_log = torch.log(torch.clamp(y_true, 0.0001))
 97 |         d = first_log - second_log
 98 |         loss = torch.mean(d * d) - self.lamada * torch.mean(d) * torch.mean(d)
 99 |         return loss
100 | 
101 | def create_depth_color(depth):
102 |     d_min = np.min(depth)
103 |     d_max = np.max(depth)
104 |     depth_relative = (depth - d_min) / (d_max - d_min)
105 |     depth = (255 * cmap(depth_relative)[:, :, :3])
106 |     return depth
107 | 
108 | 
109 | def save_image(model, x, y, batch, mode="train"):
110 |     pred = model(x)
111 |     npimg = pred.cpu().detach().numpy()
112 |     depth = create_depth_color(np.transpose(npimg[0], [1,2,0])[:, :, 0])
113 |     target = create_depth_color(np.transpose(y[0].cpu().numpy(), [1,2,0])[:, :, 0])
114 |     orig = 255 * np.transpose(x[0].cpu().numpy(), [1,2,0])
115 | 
116 |     img = np.concatenate((orig, target, depth), axis =1)
117 |     
118 |     img = Image.fromarray(img.astype('uint8'))
119 |     img.save('saved_images/%s_image_%d.jpg'%(mode, batch))
120 |     
121 | def adjust_learning_rate(optimizer, epoch, lr_init):
122 |     """Sets the learning rate to the initial LR decayed by 2 every 5 epochs"""
123 |     lr = lr_init * (0.5 ** (epoch // 5))
124 |     for param_group in optimizer.param_groups:
125 |         param_group['lr'] = lr
126 | 
127 | def train(train_loader, val_loader, model, criterion_L1, criterion_MSE, 
128 |           criterion_berHu, optimizer, epoch, batch_size):
129 |     
130 |     model.train()  # switch to train mode
131 |     eval_mode = False
132 |     init_lr = optimizer.param_groups[0]['lr']
133 |     
134 |     for iter_ in range(num_batches):
135 |         input, target = next(train_loader.get_one_batch(batch_size))
136 |         input, target = input, target
137 |         input, target = input.cuda(), target.cuda()
138 |         torch.cuda.synchronize()
139 |         
140 |         optimizer.zero_grad()
141 |         
142 |         pred = model(input)
143 |         
144 |         loss_L1 = criterion_L1(pred, target)
145 |         loss_MSE = criterion_MSE(pred, target)
146 |         loss_berHu = criterion_berHu(pred, target)
147 |         loss_SI = criterion_SI(pred, target)
148 | 
149 |         writer.add_scalar('L1 Loss', loss_L1.item(), 25*epoch + iter_ + 1)
150 |         writer.add_scalar('MSE Loss', loss_MSE.item(), 25*epoch + iter_ + 1)
151 |         writer.add_scalar('SI Loss', loss_SI.item(), 25*epoch + iter_ + 1)
152 |         writer.add_scalar('berHu Loss', loss_berHu.item(), 25*epoch + iter_ + 1)
153 | 
154 |         writer.add_scalars('loss/metrics', { 
155 |                               "L1": loss_L1.item(), "MSE": loss_MSE.item(), 
156 |                               "SI": loss_SI.item(), "berHu": loss_berHu.item()}
157 |                               , 25 * epoch + iter_)
158 |         loss_gen = loss_SI + 5 * loss_L1
159 |         loss_gen.backward()
160 |         optimizer.step()
161 |         
162 |         if (iter_ + 1) % 10 == 0:
163 |             save_image(model, input, target, iter_)
164 |             print('Train Epoch: {} Batch: [{}/{}], SI: {:0.4f}, L1 ={:0.3f}, MSE={:0.3f}, berHu={:0.3f}'.format(
165 |                 epoch, iter_ + 1, num_batches, loss_SI.item(),
166 |                 loss_L1.item(), loss_MSE.item(), loss_berHu.item()))
167 | 
168 | 
169 | train_loader = DataLoader("../cnn_depth_tensorflow/data/nyu_datasets/")
170 | val_loader = DataLoader("../cnn_depth_tensorflow/data/nyu_datasets/", mode="val")
171 | 
172 | model = UNet(3, 1).double()
173 | model = nn.DataParallel(model).cuda()
174 | 
175 | optimizer = torch.optim.Adam(model.parameters(), lr = 0.0002, weight_decay=1e-4)
176 | 
177 | criterion_L1 = MaskedL1Loss()
178 | criterion_berHu = berHuLoss()
179 | criterion_MSE = MaskedMSELoss()
180 | criterion_SI = ScaleInvariantError()
181 | 
182 | batch_size = 8
183 | num_epochs = 40
184 | num_batches = len(train_loader)//batch_size
185 | 
186 | for epoch in range(25):
187 |     adjust_learning_rate(optimizer, epoch, 0.001)
188 |     train(train_loader, val_loader, model, criterion_L1, criterion_MSE, criterion_berHu, 
189 |           optimizer, epoch, batch_size)
190 |      
191 |     torch.save({
192 |             'epoch': epoch,
193 |             'model_state_dict': model.state_dict(),
194 |             'optimizer_state_dict': optimizer.state_dict(),
195 |             }, './unet_si.pth')
196 | 
197 | 
198 | 


--------------------------------------------------------------------------------
/unet.py:
--------------------------------------------------------------------------------
  1 | import torch
  2 | import torch.nn as nn
  3 | import torch.nn.functional as F
  4 | 
  5 | 
  6 | class double_conv(nn.Module):
  7 |     '''(conv => BN => ReLU) * 2'''
  8 |     def __init__(self, in_ch, out_ch):
  9 |         super(double_conv, self).__init__()
 10 |         self.conv = nn.Sequential(
 11 |             nn.Conv2d(in_ch, out_ch, 3, padding=1),
 12 |             nn.BatchNorm2d(out_ch),
 13 |             nn.ReLU(inplace=True),
 14 |             nn.Conv2d(out_ch, out_ch, 3, padding=1),
 15 |             nn.BatchNorm2d(out_ch),
 16 |             nn.ReLU(inplace=True)
 17 |         )
 18 | 
 19 |     def forward(self, x):
 20 |         x = self.conv(x)
 21 |         return x
 22 | 
 23 | 
 24 | class inconv(nn.Module):
 25 |     def __init__(self, in_ch, out_ch):
 26 |         super(inconv, self).__init__()
 27 |         self.conv = double_conv(in_ch, out_ch)
 28 | 
 29 |     def forward(self, x):
 30 |         x = self.conv(x)
 31 |         return x
 32 | 
 33 | 
 34 | class down(nn.Module):
 35 |     def __init__(self, in_ch, out_ch):
 36 |         super(down, self).__init__()
 37 |         self.mpconv = nn.Sequential(
 38 |             nn.MaxPool2d(2),
 39 |             double_conv(in_ch, out_ch)
 40 |         )
 41 | 
 42 |     def forward(self, x):
 43 |         x = self.mpconv(x)
 44 |         return x
 45 | 
 46 | 
 47 | class up(nn.Module):
 48 |     def __init__(self, in_ch, out_ch, bilinear=True):
 49 |         super(up, self).__init__()
 50 | 
 51 |         #  would be a nice idea if the upsampling could be learned too,
 52 |         #  but my machine do not have enough memory to handle all those weights
 53 |         if bilinear:
 54 |             self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
 55 |         else:
 56 |             self.up = nn.ConvTranspose2d(in_ch//2, in_ch//2, 2, stride=2)
 57 | 
 58 |         self.conv = double_conv(in_ch, out_ch)
 59 | 
 60 |     def forward(self, x1, x2):
 61 |         x1 = self.up(x1)
 62 |         
 63 |         # input is CHW
 64 |         diffY = x2.size()[2] - x1.size()[2]
 65 |         diffX = x2.size()[3] - x1.size()[3]
 66 | 
 67 |         x1 = F.pad(x1, (diffX // 2, diffX - diffX//2,
 68 |                         diffY // 2, diffY - diffY//2))
 69 |         
 70 |         # for padding issues, see 
 71 |         # https://github.com/HaiyongJiang/U-Net-Pytorch-Unstructured-Buggy/commit/0e854509c2cea854e247a9c615f175f76fbb2e3a
 72 |         # https://github.com/xiaopeng-liao/Pytorch-UNet/commit/8ebac70e633bac59fc22bb5195e513d5832fb3bd
 73 | 
 74 |         x = torch.cat([x2, x1], dim=1)
 75 |         x = self.conv(x)
 76 |         return x
 77 | 
 78 | 
 79 | class outconv(nn.Module):
 80 |     def __init__(self, in_ch, out_ch):
 81 |         super(outconv, self).__init__()
 82 |         self.conv = nn.Conv2d(in_ch, out_ch, 1)
 83 | 
 84 |     def forward(self, x):
 85 |         x = self.conv(x)
 86 |         return x
 87 |     
 88 | class UNet(nn.Module):
 89 |     def __init__(self, n_channels, n_classes):
 90 |         super(UNet, self).__init__()
 91 |         self.inc = inconv(n_channels, 64)
 92 |         self.down1 = down(64, 128)
 93 |         self.down2 = down(128, 256)
 94 |         self.down3 = down(256, 512)
 95 |         self.down4 = down(512, 512)
 96 |         self.up1 = up(1024, 256)
 97 |         self.up2 = up(512, 128)
 98 |         self.up3 = up(256, 64)
 99 |         self.up4 = up(128, 64)
100 |         self.outc = outconv(64, n_classes)
101 | 
102 |     def forward(self, x):
103 |         x1 = self.inc(x)
104 |         x2 = self.down1(x1)
105 |         x3 = self.down2(x2)
106 |         x4 = self.down3(x3)
107 |         x5 = self.down4(x4)
108 |         x = self.up1(x5, x4)
109 |         x = self.up2(x, x3)
110 |         x = self.up3(x, x2)
111 |         x = self.up4(x, x1)
112 |         x = self.outc(x)
113 |         return F.sigmoid(x)
114 | 


--------------------------------------------------------------------------------