├── .idea ├── .gitignore ├── SLM.iml ├── dictionaries │ └── XiaoYuan.xml ├── inspectionProfiles │ ├── Project_Default.xml │ └── profiles_settings.xml ├── misc.xml ├── modules.xml └── vcs.xml ├── LICENSE ├── README.md ├── README_v2.md ├── SECURITY.md ├── evaluations └── SLM.py ├── figure ├── SeLo.jpg ├── compare.jpg ├── csm3.jpg ├── demo.gif ├── fig1.jpg ├── indicator.jpg ├── indicator_verify.jpg ├── sample.jpg ├── selo_with_subtask.jpg ├── stage1.jpg ├── stage2.jpg ├── time_analse.png └── visual_data2.jpg ├── predict ├── __pycache__ │ ├── model_init.cpython-37.pyc │ └── vocabs.cpython-37.pyc ├── cache │ └── README.md ├── checkpoints │ └── README.md ├── data │ └── rsitmd_precomp │ │ ├── test_caps.txt │ │ ├── test_filename.txt │ │ ├── train_caps.txt │ │ └── train_filename.txt ├── evaluate_selo.py ├── generate_selo.py ├── generate_selo_v2.py ├── layers │ ├── AMFMN.py │ ├── AMFMN_utils.py │ ├── LW_MCR.py │ ├── MCR_utils.py │ ├── __pycache__ │ │ ├── AMFMN.cpython-37.pyc │ │ ├── AMFMN_utils.cpython-37.pyc │ │ └── seq2vec.cpython-37.pyc │ └── seq2vec.py ├── model_encoder.py ├── model_init.py ├── option │ └── RSITMD │ │ ├── RSITMD_AMFMN.yaml │ │ └── RSITMD_LW_MCR.yaml ├── test_data │ └── sparseresidential_3814.tif ├── util │ └── convert_data.py ├── utils.py ├── vocab │ └── rsitmd_splits_vocab.json └── vocabs.py ├── requirements.txt └── test_data ├── annotations └── anno.json ├── imgs └── README.md └── visual_testdata.py /.idea/.gitignore: -------------------------------------------------------------------------------- 1 | # Default ignored files 2 | /shelf/ 3 | /workspace.xml 4 | -------------------------------------------------------------------------------- /.idea/SLM.iml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/dictionaries/XiaoYuan.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | -------------------------------------------------------------------------------- /.idea/inspectionProfiles/Project_Default.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 33 | -------------------------------------------------------------------------------- /.idea/inspectionProfiles/profiles_settings.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 6 | -------------------------------------------------------------------------------- /.idea/misc.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | -------------------------------------------------------------------------------- /.idea/modules.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/vcs.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2021 Govern 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## The offical code for paper ["Learning to Evaluate Performance of Multi-modal Semantic Localization", TGRS 2022.](https://arxiv.org/abs/2209.06515) 2 | ##### Author: Zhiqiang Yuan, Chongyang Li, Zhuoying Pan, et. al 3 | 4 | 5 | ![Supported Python versions](https://img.shields.io/badge/python-3.7-blue.svg) 6 | ![Supported OS](https://img.shields.io/badge/Supported%20OS-Linux-yellow.svg) 7 | ![npm License](https://img.shields.io/npm/l/mithril.svg) 8 | 9 | 10 | ### ------------------------------------------------------------------------------------- 11 | ### Welcome :+1:_`Fork and Star`_:+1:, then we'll let you know when we update 12 | 13 | ### ------------------------------------------------------------------------------------- 14 | ### We recently released SeLo v2[\[link\]](https://github.com/xiaoyuan1996/SemanticLocalizationMetrics/blob/master/README_v2.md)., which improves SeLo from the speed and accuracy. 15 | 16 | ### ------------------------------------------------------------------------------------- 17 | 18 | #### Contexts 19 | 20 | * [Introduction](#introduction) 21 | * [Dataset And Metrics](#dataset-and-metrics) 22 | * [Testdata](#testdata) 23 | * [Metrics](#metrics) 24 | * [Baselines](#baselines) 25 | * [Comparison of SeLo Performance on Different Trainsets](#comparison-of-selo-performance-on-different-trainsets) 26 | * [Comparison of SeLo Performance on Different Scales](#comparison-of-selo-performance-on-different-scales) 27 | * [Comparison of SeLo Performance on Different Retrieval Models](#comparison-of-selo-performance-on-different-retrieval-models) 28 | * [Analysis of Time Consumption](#analysis-of-time-consumption) 29 | * [Implementation](#implementation) 30 | * [Environment](#environment) 31 | * [Run The Demo](#run-the-demo) 32 | * [Customize Model](#customize-model) 33 | * [Epilogue](#epilogue) 34 | * [Citation](#citation) 35 | ### ------------------------------------------------------------------------------------- 36 | 37 | 38 | 39 | ## INTRODUCTION 40 | An official evaluation metric for semantic localization. 41 | 42 | compare 43 | 44 | **Fig.1.** (a) Results of airplane detection. (b) Results of semantic localization with query of ``white planes parked in the open space of the white airport''. Compared with tasks such as detection, SeLo achieves semantic-level retrieval with only caption-level annotation during training, which can adapt to higher-level retrieval tasks. 45 | 46 | 47 | 48 | shwon 49 | 50 | **Fig.2.** Visualization of SeLo with query of "the red rails where the grey train is located run through the residential area". 51 | 52 | 53 | The semantic localization (SeLo) task refers to using cross-modal information such as text to quickly localize RS images at the semantic level [\[link\]](https://ieeexplore.ieee.org/document/9437331). 54 | This task implements semantic-level detection, which only uses caption-level supervision information. 55 | In our opinion, it is a meaningful and interesting work, which realizes the unification of sub-tasks such as detection and segmentation. 56 | 57 | ![visual image](./figure/SeLo.jpg) 58 | 59 | **Fig.3.** Framework of Semantic Localization. After multi-scale segmentation of large RS images, we perform cross-modal similarity calculation on query and multiple slices. The calculated regional probabilities are then utilized to pixel-level averaging, which generates the SeLo map after further noise suppression. 60 | 61 | We contribute test sets, evaluation metrics and baselines for semantic localization, and provide a detailed demo to use this evaluation framework. 62 | Any questions can open a Github [issue](https://github.com/xiaoyuan1996/SemanticLocalizationMetrics/issues). 63 | Start and enjoy! 64 | 65 | 66 | 67 | ### ------------------------------------------------------------------------------------- 68 | 69 | ## DATASET AND METRICS 70 | 71 | #### TESTDATA 72 | 73 | We contribute a semantic localization testset to provide systematic evaluation for SeLo task. 74 | The images in SLT come from Google Earth, and Fig. 4 exhibits several samples from the testset. 75 | Every sample includes a large image in RS scene with the size of 3k × 2k to 10k × 10k, a query sentence, and one or more corresponding semantic bounding boxes. 76 | 77 | sample 78 | 79 | 80 | **Fig.4.** Four samples of Semantic Localization Testset. (a) Query: “ships without cargo floating on the black sea are docked in the port”. (b) Query: “a white 81 | airplane ready to take off on a grayblack runway”. (c) Query: “some cars are parked in a parking lot surrounded by green woods”. (d) Query: “the green 82 | football field is surrounded by a red track”. 83 | 84 | **TABLE I** Quantitative Statistics of Semantic Localization Testset. 85 | 86 | | Parameter | Value | Parameter | Value | 87 | | ------------- | ------| ---------------------| -------| 88 | | Word Number | 160 | Caption Ave Length | 11.2 | 89 | | Sample Number | 59 | Ave Resolution Ratio (m) | 0.3245 | 90 | | Channel Number| 3 | Ave Region Number | 1.75 | 91 | | Image Number | 22 | Ave Attention Ratio | 0.068 | 92 | 93 | 94 | 95 | 96 | #### METRICS 97 | 98 | We systematically model and study semantic localization in detail, and propose multiple discriminative evaluation metrics to quantify this task based on significant area proportion, attention shift distance, and discrete attention distance. 99 | 100 | shwon 101 | 102 | **Fig.5.** Three proposed evaluation metrics for semantic localization. (a) Rsu aims to calculate the attention ratio of the ground-truth area to the useless 103 | area. (b) Ras attempts to quantify the shift distance of the attention from the GT center. (c) Rda evaluates the discreteness of the generated attention from 104 | probability divergence distance and candidate attention number. 105 | 106 | **TABLE II** Explanation of the indicators. 107 | 108 | | Indicator | Range | Meaning | 109 | | --------- | ------| ---------| 110 | | Rsu | ↑ [ 0 ~ 1 ] | Calc the salient area proportion | 111 | | Ras | ↓ [ 0 ~ 1 ] | Makes attention center close to annotation center | 112 | | Rda | ↑ [ 0 ~ 1 ] | Makes attention center focus on one point | 113 | | Rmi | ↑ [ 0 ~ 1 ] | Calculate the mean indicator of SeLo task | 114 | 115 | 116 | 117 | shwon 118 | 119 | **Fig.6.** Qualitative analysis of SeLo indicator validity. (a) Query: “eight large white oil storage tanks built on grey concrete floor”. (b) Query: “a white plane 120 | parked in a tawny clearing inside the airport”. (c) Query: “lots of white and black planes parked inside the grey and white airport”. 121 | 122 | ## BASELINES 123 | 124 | All experiments all carried out at Intel(R) Xeon(R) Gold 6226R CPU @2.90GHz and a single NVIDIA RTX 3090 GPU. 125 | 126 | #### Comparison of SeLo Performance on Different Trainsets 127 | 128 | | Trainset | ↑ Rsu | ↑ Rda | ↓ Ras | ↑ Rmi | 129 | | ------------- | ------| -------------- | -------| -------| 130 | | Sydney | 0.5844 | 0.5670 | 0.5026 | 0.5496 | 131 | | UCM | 0.5821 | 0.4715 | 0.5277 | 0.5160 | 132 | | RSITMD| **0.6920** | **0.6667** | **0.3323** | **0.6772** | 133 | | RSICD | 0.6661 | 0.5773 | 0.3875 | 0.6251 134 | 135 | 136 | #### Comparison of SeLo Performance on Different Scales 137 | 138 | | | Scale-128 | Scale-256 | Scale-512 | Scale-768 | ↑ Rsu | ↑ Rda | ↓ Ras | ↑ Rmi | Time (m) 139 | | --- | ----------| ----------| ----------| -----------| ----- | ----- | ----- | ----- | ----- | 140 | | s1 | √ | √ | | | 0.6389 | 0.6488 | 0.2878 | 0.6670 | 33.81 | 141 | | s2 | | √ | √ | | 0.6839 | 0.6030 | 0.3326 | 0.6579 | 14.25 | 142 | | s3 | | | √ | √ | 0.6897 | 0.6371 | 0.3933 | 0.6475 | **11.23** | 143 | | s4 | √ | √ | √ | | 0.6682 | **0.7072** | **0.2694** | **0.6998** | 34.60 | 144 | | s5 | | √ | √ | √ | **0.6920** | 0.6667 | 0.3323 | 0.6772 | 16.92 | 145 | | s6 | √ | √ | √ | √ | 0.6809 | 0.6884 | 0.3025 | 0.6886 | 36.28 | 146 | 147 | 148 | #### Comparison of SeLo Performance on Different Retrieval Models 149 | 150 | | Trainset | ↑ Rsu | ↑ Rda | ↓ Ras | ↑ Rmi | Time (m) | 151 | | ------------- | ------| -------------- | -------| -------| -------- | 152 | | VSE++ | 0.6364 | 0.5829 | 0.4166 | 0.6045 | 15.61 153 | | LW-MCR | 0.6698 | 0.6021 | 0.4335 | 0.6167 | **15.47** 154 | | SCAN| 0.6421 | 0.6132 | 0.3871 | 0.6247 | 16.32 155 | | CAMP | 0.6819 | 0.6314 | 0.3912 | 0.6437 | 18.24 156 | | AMFMN | **0.6920** | **0.6667** | **0.3323** | **0.6772** | 16.92 157 | 158 | #### Analysis of Time Consumption 159 | 160 | | Scale (128, 256) | Cut | Sim | Gnt | Flt | Total | 161 | | ------------- | ------| -------------- | -------| -------| ------| 162 | | Times(m) | 2.85| 20.60 | 7.40| 0.73| 33.81| 163 | | Rate(%) | 8.42| 60.94 | 21.88| 2.16| -| 164 | 165 | | Scale (512, 768) | Cut | Sim | Gnt | Flt | Total | 166 | | ------------- | ------| -------------- | -------| -------| ------| 167 | | Times(m) | 0.46| 1.17 | 6.96| 0.67| 11.23| 168 | | Rate(%) | 4.06| 10.42 | 61.98| 5.97| -| 169 | 170 | | Scale (256, 512, 768) | Cut | Sim | Gnt | Flt | Total | 171 | | ------------- | ------| -------------- | -------| -------| ------| 172 | | Times(m) | 0.93| 5.72 | 7.38| 0.74| 16.92| 173 | | Rate(%) | 5.52| 33.82 | 43.60| 4.37| -| 174 | 175 | ## IMPLEMENTATION 176 | 177 | #### ENVIRONMENT 178 | 179 | 1.Pull our project and install the requirements, make sure the code path only include English: 180 | 181 | ``` 182 | $ apt-get install python3 183 | $ git clone git@github.com:xiaoyuan1996/SemanticLocalizationMetrics.git 184 | $ cd SemanticLocalizationMetrics 185 | $ pip install -r requirements.txt 186 | ``` 187 | 188 | 2.Prepare checkpoints and test iamges: 189 | 190 | 191 | * Download pretrain checkpoints **AMFMN.tar** to **./predict/checkpoints/** from [GoogleDriver](https://drive.google.com/drive/folders/1LISJHiLVxPCiry1i7xJtOvuQ77nbEZD1?usp=sharing), make sure: 192 | 193 | + ./predict/checkpoints/ 194 | + xxx.pth 195 | 196 | * Download test images **SLM_testimgs.zip** to **./test_data/imgs/** from [BaiduYun (passwd: NIST)](https://pan.baidu.com/s/1DyRbY7s3jx-ZCWbcC_XHlw) or [GoogleDriver](https://drive.google.com/drive/folders/1LISJHiLVxPCiry1i7xJtOvuQ77nbEZD1?usp=sharing), make sure: 197 | 198 | + ./test_data/imgs/ 199 | + xxx.tif 200 | 201 | 3.Download SkipThought Files to **/data** from [seq2vec (Password:NIST)](https://pan.baidu.com/s/1FOPldSGO6ctETiXMlPGC8g) (or other path, but you should change the **option['model']['seq2vec']['dir_st']**) 202 | 203 | 4.Check the environments 204 | 205 | ``` 206 | $ cd predict 207 | $ python model_encoder.py 208 | 209 | visual_vector: (512,) 210 | text_vector: (512,) 211 | Encoder test successful! 212 | Calc sim successful! 213 | ``` 214 | 215 | 216 | #### RUN THE DEMO 217 | Run the follow command, and generated SeLo maps will be saved in **cache/**. 218 | ``` 219 | $ cd predict 220 | $ nohup python generate_selo.py --cache_path cache/RSITMD_AMFMN 221 | $ tail -f cache/RSITMD_AMFMN/log.txt 222 | 223 | 2022-05-05 22:01:58,180 - __main__ - INFO - Processing 31/59: 20.jpg 224 | 2022-05-05 22:01:58,180 - __main__ - INFO - Corresponding text: lots of white and black planes parked inside the grey and white airport. 225 | 226 | 2022-05-05 22:01:59,518 - __main__ - INFO - img size:10000x10001 227 | 2022-05-05 22:01:59,518 - __main__ - INFO - Start split images with step 256 228 | 2022-05-05 22:02:02,657 - __main__ - INFO - Start split images with step 512 229 | 2022-05-05 22:02:04,077 - __main__ - INFO - Start split images with step 768 230 | 2022-05-05 22:02:04,818 - __main__ - INFO - Image ../test_data/imgs/20.jpg has been split successfully. 231 | 2022-05-05 22:02:04,819 - __main__ - INFO - Start calculate similarities ... 232 | 2022-05-05 22:02:48,182 - __main__ - INFO - Calculate similarities in 43.36335849761963s 233 | 2022-05-05 22:02:48,182 - __main__ - INFO - Start generate heatmap ... 234 | 2022-05-05 22:03:40,673 - __main__ - INFO - Generate finished, start optim ... 235 | 2022-05-05 22:03:45,500 - __main__ - INFO - Generate heatmap in 57.31790471076965s 236 | 2022-05-05 22:03:45,500 - __main__ - INFO - Saving heatmap in cache/heatmap_31.jpg ... 237 | 2022-05-05 22:03:45,501 - __main__ - INFO - Saving heatmap in cache/addmap_31.jpg ... 238 | 2022-05-05 22:03:45,501 - __main__ - INFO - Saving heatmap in cache/probmap_31.jpg ... 239 | 2022-05-05 22:03:48,540 - __main__ - INFO - Saved ok. 240 | 2022-05-05 22:03:59,562 - root - INFO - Eval cache/probmap_31.jpg 241 | 2022-05-05 22:03:59,562 - root - INFO - +++++++ Calc the SLM METRICS +++++++ 242 | 2022-05-05 22:03:59,562 - root - INFO - ++++ ↑ Rsu [0 ~ 1]:0.9281 ++++ 243 | 2022-05-05 22:03:59,562 - root - INFO - ++++ ↑ Rda [0 ~ 1]:0.4689 ++++ 244 | 2022-05-05 22:03:59,562 - root - INFO - ++++ ↓ Ras [0 ~ 1]:0.0633 ++++ 245 | 2022-05-05 22:03:59,562 - root - INFO - ++++ ↑ Rmi [0 ~ 1]:0.8163 ++++ 246 | 2022-05-05 22:03:59,562 - root - INFO - ++++++++++++++++++++++++++++++++++++ 247 | ... 248 | 249 | $ ls cache/RSITMD_AMFMN 250 | ``` 251 | 252 | 253 | 254 | #### CUSTOMIZE MODEL 255 | 256 | 1. Put the pretrain ckpt file to **checkpoints**. 257 | 2. Add your own model to **layers** and corresponding config yaml to **options/**. 258 | 3. Change **model_init.model_init** to your own models. 259 | 4. Add the class of **EncoderSpecModel** to **model_encoder.py**. 260 | 5. Run: 261 | ``` 262 | python generate_selo.py --yaml_path option/xxx.yaml --cache_path cache/xxx 263 | ``` 264 | 265 | ## EPILOGUE 266 | So far, our attitude towards the semantic localization task is positive and optimistic, which realizes the detection at the semantic level with only the annotation at the caption level. 267 | We sincerely hope that this project will facilitate the development of semantic localization tasks. 268 | We welcome researchers to look into this direction, which is a possibility to achieve refined object semantic detection. 269 | 270 | ![visual image](./figure/selo_with_subtask.jpg) 271 | 272 | **Fig.7.** Combine SeLo with other tasks. The top of the figure shows the detection results after add the SeLo map with query of “two parallel green 273 | playgrounds”. The bottom of the figure shows the road extraction results after add the SeLo map with query of “the red rails where the grey train is located 274 | run through the residential area”. (a) Source images. (b) Results of specific tasks. (c) Results of specific SeLo maps. (d) Fusion results of specific tasks and 275 | SeLo map. 276 | 277 | ## CITATION 278 | ``` 279 | Z. Yuan et al., "Learning to Evaluate Performance of Multi-modal Semantic Localization," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2022.3207171. 280 | ``` 281 | 282 | ## OTHER CITATION 283 | ``` 284 | Z. Yuan et al., "Exploring a Fine-Grained Multiscale Method for Cross-Modal Remote Sensing Image Retrieval," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2021.3078451. 285 | 286 | Z. Yuan et al., "A Lightweight Multi-scale Crossmodal Text-Image Retrieval Method In Remote Sensing," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2021.3124252. 287 | 288 | Z. Yuan et al., "Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2022.3163706. 289 | ``` 290 | -------------------------------------------------------------------------------- /README_v2.md: -------------------------------------------------------------------------------- 1 | # Introduction 2 | 3 | Based on SeLo, we conduct a more in-depth exploration and propose SeLo v2. Compared with initial SeLo framework, SeLo v2 has better semantic localization performance with only half the time consumed, which show the great advantages of SeLo v2 in terms of time and accuracy. 4 | 5 | ![fig1](./figure/fig1.jpg) 6 | 7 | (a) Caption (b) SeLo (c) SeLo v2. SeLo refers to the task of obtaining the most relevant locations in large-scale RS images using semantic information such as text. Compared with SeLo, SeLo v2 weakens the attention of non-target pixels and achieves more detailed semantic-level retrieval. 8 | 9 | # Method 10 | 11 | ## Multilevel Likelihood Expansion 12 | 13 | ![stage1](./figure/stage1.jpg) 14 | 15 | In the first stage, the RS image is cropped with a larger scale and a single ratio. The purpose of this cropping method is to make the large slice contain more semantic information and roughly lock the slice where the target object may be located. 16 | 17 | ![stage2](./figure/stage2.jpg) 18 | 19 | In the second stage, we use the pixel continuity of the target object to perform multi-scale expansion inward and outward on the initially determined large slice. 20 | 21 | ## Chaotic Self-feeding Mechanism 22 | 23 | ![csm](./figure/csm3.jpg) 24 | 25 | Fig. 3. Schematic diagram of the proposed chaotic self-feeding mechanism.In the case of a small number of samples, the features generated by the text encoder are not stable. For this reason, CSM regards it as a high-frequency feature, and generates a fundamental frequency representationthrough confusion to obtain a more stable representation. 26 | 27 | # Results 28 | 29 | ![visual_data](./figure/visual_data2.jpg) 30 | 31 | Quantitative Comparison of SeLo Performance on Different 32 | 33 | | Model | ↑Rsu | ↑Rda | ↓Ras | ↑Rmi | Time(m) | 34 | | :------------: | :--------: | :--------: | :--------: | :--------: | :-------: | 35 | | VSE++ | 0.6364 | 0.5829 | 0.4166 | 0.6045 | 25.23 | 36 | | LW-MCR | 0.6698 | 0.6021 | 0.4335 | 0.6167 | 25.01 | 37 | | SCAN | 0.6421 | 0.6132 | 0.3871 | 0.6247 | 26.38 | 38 | | CAMP | 0.6819 | 0.6314 | 0.3912 | 0.6437 | 29.48 | 39 | | AMFMN | 0.6920 | 0.6667 | 0.3323 | 0.6772 | 27.35 | 40 | | **AMFMN-ours** | **0.7199** | **0.6658** | **0.2925** | **0.7021** | **13.34** | 41 | 42 | Experimental Results of SeLo V2 Method with Different Structures 43 | 44 | | Ablation Method | VS | MLE | STE | CSM | Med | Gass | ↑Rsu | ↑Rda | ↓Ras | ↑Rmi | Time(m) | 45 | | :-------------: | ---- | :--: | ---- | ---- | ---- | :--: | ------ | ------ | ------ | ------ | ------- | 46 | | m1 | √ | | √ | | √ | | 0.6920 | 0.6667 | 0.3323 | 0.6772 | 27.3510 | 47 | | m2 | | √ | √ | | √ | | 0.6976 | 0.6381 | 0.3302 | 0.6730 | 22.7770 | 48 | | m3 | √ | | √ | | | √ | 0.6942 | 0.6792 | 0.3164 | 0.6867 | 23.2008 | 49 | | m4 | | √ | √ | | | √ | 0.7002 | 0.6575 | 0.3330 | 0.6780 | 14.3155 | 50 | | m5 | | √ | | √ | | √ | 0.7199 | 0.6658 | 0.2975 | 0.7021 | 13.3447 | 51 | 52 | This paper proposes a multilevel likelihood expansion and chaotic self-feeding mechanism for faster and higher SeLo task. Qualitative and quantitative experiments show that the proposed SeLo v2 has effectively promoted the development of RS image SeLo tasks. 53 | -------------------------------------------------------------------------------- /SECURITY.md: -------------------------------------------------------------------------------- 1 | # Security Policy 2 | 3 | ## Supported Versions 4 | 5 | Use this section to tell people about which versions of your project are 6 | currently being supported with security updates. 7 | 8 | | Version | Supported | 9 | | ------- | ------------------ | 10 | | 5.1.x | :white_check_mark: | 11 | | 5.0.x | :x: | 12 | | 4.0.x | :white_check_mark: | 13 | | < 4.0 | :x: | 14 | 15 | ## Reporting a Vulnerability 16 | 17 | Use this section to tell people how to report a vulnerability. 18 | 19 | Tell them where to go, how often they can expect to get an update on a 20 | reported vulnerability, what to expect if the vulnerability is accepted or 21 | declined, etc. 22 | -------------------------------------------------------------------------------- /evaluations/SLM.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/03/08 6 | 7 | import logging 8 | from functools import reduce 9 | 10 | import cv2 11 | import numpy as np 12 | from scipy.ndimage import maximum_filter 13 | from skimage import measure 14 | 15 | 16 | class SLM(object): 17 | def __init__(self): 18 | # logging 19 | logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s') 20 | self.logger = logging.getLogger() 21 | 22 | # parameters 23 | self.rsu_beta = 0.707 24 | self.rsu_eps = 1e-7 25 | 26 | self.ras_expand_factor = 1.5 27 | self.ras_filter_times = 5 28 | self.ras_scala_beta = 3 29 | 30 | self.rda_eta = 0.5 31 | 32 | self.rmi_wsu = 0.4 33 | self.rmi_was = 0.35 34 | self.rmi_wda = 0.25 35 | 36 | # visual settings 37 | self.visual_ras = False 38 | self.src_addmap_path = None 39 | 40 | # sum indicator 41 | self.all_metrics = self._format_output_dict() 42 | 43 | def _format_output_dict(self, *params): 44 | """ 45 | format output dict 46 | :param params: keys 47 | :return: format dict 48 | """ 49 | len_params = len(params) 50 | if len_params == 0: init_param = [[] for i in range(4)] 51 | elif len_params == 4: init_param = params 52 | else: raise NotImplementedError 53 | 54 | return { 55 | "↑ Rsu [0 ~ 1]": init_param[0], 56 | "↑ Rda [0 ~ 1]": init_param[1], 57 | "↓ Ras [0 ~ 1]": init_param[2], 58 | "↑ Rmi [0 ~ 1]": init_param[3] 59 | } 60 | 61 | def logging_acc(self, metrics_dict, prob_path = None, ave = False): 62 | """ 63 | logging the metrics 64 | :param metrics_dict: dict of metrics 65 | :param prob_path: path 66 | :return: 0 67 | """ 68 | 69 | if not ave: 70 | self.logger.info("Eval {}".format(prob_path)) 71 | else: 72 | self.logger.info("+++++++++++++++Average++++++++++++++") 73 | 74 | self.logger.info("+++++++ Calc the SLM METRICS +++++++") 75 | for metric, value in metrics_dict.items(): 76 | self.logger.info("++++ {}:{:.4f} ++++".format(metric, value)) 77 | self.logger.info("++++++++++++++++++++++++++++++++++++\n") 78 | 79 | def set_visual_options(self, visual_ras, src_addmap_path): 80 | """ 81 | set visual options 82 | :param visual_ras: flag 83 | :param src_addmap_path: set src addmap path 84 | """ 85 | self.visual_ras = visual_ras 86 | self.src_addmap_path = src_addmap_path 87 | return True 88 | 89 | def read_gray_to_prob(self, probmap_path): 90 | """ 91 | Read the prob maps, and trans to probility 92 | :param probmap_path: probmap routh 93 | :return: probability 94 | """ 95 | gray_image = cv2.imread(probmap_path, cv2.IMREAD_GRAYSCALE) 96 | prob = gray_image / 255.0 97 | return prob 98 | 99 | def generate_mask_by_points(self, prob, points_list): 100 | """ 101 | Generate mask by regions 102 | :param prob: probability 103 | :param points_list: regions 104 | :return: mask 105 | """ 106 | H, W = prob.shape 107 | 108 | mask = np.zeros((H, W)) 109 | 110 | points_list = [np.array(i, np.int32) for i in points_list] 111 | 112 | # fill 113 | cv2.fillPoly(mask, points_list, 1) 114 | 115 | return mask 116 | 117 | def _get_region_center_radius(self, region_point): 118 | """ 119 | get the region center and radius 120 | :param region_point: regions 121 | :return: mid_x, mid_y, radius 122 | """ 123 | mid_x = int(reduce(lambda x, y: x+y, np.array(region_point)[:, 0]) / len(region_point)) 124 | mid_y = int(reduce(lambda x, y: x+y, np.array(region_point)[:, 1]) / len(region_point)) 125 | radius = int(np.mean([np.linalg.norm(np.array(point) - np.array([mid_x, mid_y])) for point in region_point]) * self.ras_expand_factor) 126 | return mid_x, mid_y, radius 127 | 128 | def _get_prob_center_in_gray(self, prob): 129 | """ 130 | get the top point with the highest probability from the probability map 131 | :param prob: probability 132 | :return: centers 133 | """ 134 | 135 | # recover the prob 136 | gray_img = np.asarray(prob * 255.0, dtype=np.uint8) 137 | 138 | # construct continuous area 139 | continuous_area = np.asarray(gray_img > 150, np.uint8) * 255 140 | continuous_area = np.uint8(measure.label(continuous_area, connectivity=2)) 141 | 142 | # soften 143 | for i in range(self.ras_filter_times): 144 | gray_img = cv2.boxFilter(gray_img, ddepth=-1, ksize=(50, 50)) 145 | 146 | # get probability binary map 147 | mx = maximum_filter(gray_img, size=1000) 148 | gray_img = np.where(mx == gray_img, gray_img, 0) 149 | gray_img = np.asarray(gray_img > 0, np.uint8) * 255 150 | 151 | # get probability area information 152 | labels = measure.label(gray_img, connectivity=2) 153 | all_region_infos = measure.regionprops(labels) 154 | centers = [[int(i) for i in prop.centroid][::-1] for prop in all_region_infos] 155 | 156 | # construct v-center list and sort 157 | v_center = [[c[0], c[1], prob[c[1]][c[0]]] for c in centers] 158 | v_center.sort(key= lambda x: x[2], reverse=True) 159 | centers = list(map(lambda x: x[:2], v_center)) 160 | 161 | # filter centers 162 | centers = [i for i in centers if prob[i[1]][i[0]] >= 0.5] 163 | 164 | return centers, continuous_area 165 | 166 | def _get_offset_between_real_and_synthetic(self, real_center_radius, prob_centers, bina_img): 167 | """ 168 | calculate true center offset from result center 169 | :param real_center_radius: real_center_radius 170 | :param prob_centers: prob_centers 171 | :return: offsets 172 | """ 173 | 174 | # check prob_centers is not None 175 | if len(prob_centers) == 0 : return [real_center_radius[0][2]] 176 | 177 | offsets = [] 178 | for center_radius in real_center_radius: 179 | x, y, r = center_radius 180 | 181 | # calc the l2 dis 182 | dises = list(map(lambda p: np.linalg.norm(np.array([x, y] - np.array(p))), prob_centers)) 183 | 184 | # filter the dis in cicle 185 | dises = list(filter(lambda d: d <= r, dises)) 186 | 187 | # if no prob center set it to radius 188 | offsets.append(np.mean(dises) if len(dises) != 0 else r) 189 | 190 | return offsets 191 | 192 | def _trans_ras_offset_to_scalable_ras(self, offsets, centers_and_radius): 193 | """ 194 | convert distance offset to ras value 195 | :param offsets: offsets 196 | :return: centers_and_radius 197 | """ 198 | 199 | # granular transformation 200 | granular_offet = np.mean([off/v[2] for off, v in zip(offsets, centers_and_radius)]) 201 | 202 | # scala transformation 203 | granular_offet = (np.exp(self.ras_scala_beta * granular_offet) - 1) / (np.exp(self.ras_scala_beta) - 1) 204 | 205 | return granular_offet 206 | 207 | def ras(self, region_lists, prob, visual=True, src_img=None): 208 | """ 209 | calc the matric of ras: makes attention center close to annotation center 210 | :param region_lists: regions 211 | :param prob: probability 212 | :return: ras 213 | """ 214 | 215 | # get the annotation center and radius 216 | centers_and_radius = [self._get_region_center_radius(i) for i in region_lists] 217 | 218 | # get the point with the highest probability from the probability map 219 | prob_centers, bina_img = self._get_prob_center_in_gray(prob) 220 | 221 | # calculate true center offset from result center 222 | offsets = self._get_offset_between_real_and_synthetic(centers_and_radius, prob_centers, bina_img) 223 | 224 | # convert distance offset to rcs value 225 | ras = self._trans_ras_offset_to_scalable_ras(offsets, centers_and_radius) 226 | 227 | # visual 228 | if visual and (src_img != None): 229 | src_img = cv2.imread(src_img) 230 | 231 | # logging something 232 | print("centers_and_radius: ", centers_and_radius) 233 | print("prob_centers: ", prob_centers) 234 | print("offsets: ", offsets) 235 | 236 | # backup area 237 | for c_r in centers_and_radius: 238 | cv2.circle(src_img, (c_r[0], c_r[1]), c_r[2], 2, 3) 239 | 240 | # candidate points 241 | for idx, point in enumerate(prob_centers): 242 | cv2.circle(src_img, tuple(point), 6*(idx+1), 1, 4) 243 | cv2.putText(src_img, str(idx+1), tuple(point), cv2.FONT_HERSHEY_COMPLEX, 6, (0, 0, 0), 25) 244 | 245 | cv2.imwrite("./img_circle.jpg", src_img) 246 | 247 | print(prob_centers) 248 | 249 | return ras 250 | 251 | def rsu(self, prob, mask): 252 | """ 253 | calc the salient area proportion 254 | :param prob: probability 255 | :param mask: mask 256 | :return: rsu 257 | """ 258 | 259 | all_mask_value = np.sum(np.multiply(prob, mask)) 260 | all_value = np.sum(prob) 261 | H, W = np.shape(mask) 262 | all_mask = np.sum(mask) 263 | 264 | left_frac = all_mask_value / (all_value - all_mask_value + self.rsu_eps) 265 | 266 | right_frac = (H * W - all_mask) / all_mask 267 | 268 | rsu = -np.exp(-1 * self.rsu_beta * left_frac * right_frac) + 1 269 | 270 | return rsu 271 | 272 | def rda(self, region_lists, prob): 273 | """ 274 | calc the matric of rda: makes attention center focus on one point 275 | :param region_lists: regions 276 | :param prob: probability 277 | :return: rda 278 | """ 279 | 280 | # get the annotation center and radius 281 | centers_and_radius = [self._get_region_center_radius(i) for i in region_lists] 282 | 283 | # get the point with the highest probability from the probability map 284 | prob_centers, bina_img = self._get_prob_center_in_gray(prob) 285 | 286 | # set value 287 | rda = [] 288 | for c_r in centers_and_radius: 289 | x, y, r = c_r 290 | 291 | # calc the backup points 292 | backup_points = list(filter(lambda p: np.linalg.norm(np.array([x, y] - np.array(p))) <= r, prob_centers)) 293 | 294 | # margin condition 295 | len_backup_points = len(backup_points) 296 | if len_backup_points <= 1 : 297 | rda.append(float(len_backup_points)) 298 | continue 299 | 300 | # if len_backup_points >= 2, calc the attention discrete 301 | centers_attention = np.average(backup_points, axis=0) 302 | dises = list(map(lambda p: np.linalg.norm(np.array(centers_attention - np.array(p))), backup_points)) 303 | meas_dis = np.mean(dises) / r 304 | 305 | rda_single = 0.5 * (1 - meas_dis) + np.exp(- self.rda_eta * (len_backup_points + 2)) 306 | 307 | rda.append(rda_single) 308 | 309 | return np.mean(rda) 310 | 311 | 312 | def rmi(self, rsu, rda, ras): 313 | """ 314 | calculate the mean indicator 315 | :param rsu: rsu 316 | :param rda: rda 317 | :param ras: ras 318 | :return: rmi 319 | """ 320 | return self.rmi_wsu * rsu + self.rmi_was * (1 - ras) + self.rmi_wda * rda 321 | 322 | def evaluate(self, prob_path, region_list): 323 | """ 324 | evaluate the slm task 325 | :param prob_path: probability map path 326 | :param region_list: region points 327 | :return: slm metrics 328 | """ 329 | # read prob 330 | prob = self.read_gray_to_prob(prob_path) 331 | 332 | # generate mask 333 | mask = self.generate_mask_by_points(prob, region_list) 334 | 335 | # rsu 336 | rsu = self.rsu(prob, mask) 337 | 338 | # ras 339 | ras = self.ras(region_list, prob, visual=self.visual_ras, src_img=self.src_addmap_path) 340 | 341 | # rda 342 | rda = self.rda(region_list, prob) 343 | 344 | # mi 345 | rmi = self.rmi(rsu, rda, ras) 346 | 347 | # sort metrics 348 | metrics = self._format_output_dict(rsu, rda, ras, rmi) 349 | self.logging_acc(metrics, prob_path) 350 | 351 | return metrics 352 | 353 | def append_metric(self, metric): 354 | """ 355 | append metric to calc ave indicator 356 | :param metric: sort metrics 357 | """ 358 | for k in metric.keys(): 359 | self.all_metrics[k].append(metric[k]) 360 | 361 | def get_the_mean_metric(self): 362 | """ 363 | get the mean metric 364 | """ 365 | mean_metric = {} 366 | for k in self.all_metrics: 367 | mean_metric[k] = np.mean(self.all_metrics[k]) 368 | 369 | self.logging_acc(mean_metric, ave=True) 370 | return mean_metric 371 | 372 | 373 | if __name__ == "__main__": 374 | 375 | points = [ 376 | [ 377 | [ 378 | 1163.84851074219, 379 | 948.244812011719 380 | ], 381 | [ 382 | 1360.29187011719, 383 | 883.699096679688 384 | ], 385 | [ 386 | 1497.80224609375, 387 | 993.146118164063 388 | ], 389 | [ 390 | 1649.34436035156, 391 | 1299.03637695313 392 | ], 393 | [ 394 | 1660.56970214844, 395 | 2067.9716796875 396 | ], 397 | [ 398 | 1691.43933105469, 399 | 2132.51733398438 400 | ], 401 | [ 402 | 1635.31262207031, 403 | 2292.478515625 404 | ], 405 | [ 406 | 1090.88391113281, 407 | 2289.67211914063 408 | ], 409 | [ 410 | 1147.01062011719, 411 | 1481.44812011719 412 | ], 413 | [ 414 | 1278.908203125, 415 | 1374.80749511719 416 | ], 417 | [ 418 | 1166.65490722656, 419 | 1209.23376464844 420 | ], 421 | [ 422 | 1169.46118164063, 423 | 937.019470214844 424 | ], 425 | [ 426 | 1169.46118164063, 427 | 937.019470214844 428 | ], 429 | [ 430 | 1169.46118164063, 431 | 937.019470214844 432 | ] 433 | ] 434 | ] 435 | 436 | prob_path = "../predict/cache/probmap_9.jpg" 437 | add_path = "../predict/cache/addmap_9.jpg" 438 | 439 | # eval 440 | slm = SLM() 441 | # slm.set_visual_options(visual_ras=True, src_addmap_path=add_path) 442 | metrics = slm.evaluate(prob_path, region_list=points) 443 | -------------------------------------------------------------------------------- /figure/SeLo.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/SeLo.jpg -------------------------------------------------------------------------------- /figure/compare.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/compare.jpg -------------------------------------------------------------------------------- /figure/csm3.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/csm3.jpg -------------------------------------------------------------------------------- /figure/demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/demo.gif -------------------------------------------------------------------------------- /figure/fig1.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/fig1.jpg -------------------------------------------------------------------------------- /figure/indicator.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/indicator.jpg -------------------------------------------------------------------------------- /figure/indicator_verify.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/indicator_verify.jpg -------------------------------------------------------------------------------- /figure/sample.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/sample.jpg -------------------------------------------------------------------------------- /figure/selo_with_subtask.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/selo_with_subtask.jpg -------------------------------------------------------------------------------- /figure/stage1.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/stage1.jpg -------------------------------------------------------------------------------- /figure/stage2.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/stage2.jpg -------------------------------------------------------------------------------- /figure/time_analse.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/time_analse.png -------------------------------------------------------------------------------- /figure/visual_data2.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/figure/visual_data2.jpg -------------------------------------------------------------------------------- /predict/__pycache__/model_init.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/__pycache__/model_init.cpython-37.pyc -------------------------------------------------------------------------------- /predict/__pycache__/vocabs.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/__pycache__/vocabs.cpython-37.pyc -------------------------------------------------------------------------------- /predict/cache/README.md: -------------------------------------------------------------------------------- 1 | Save SeLo map here. -------------------------------------------------------------------------------- /predict/checkpoints/README.md: -------------------------------------------------------------------------------- 1 | Download SLM_checkpoints.zip, unzip it and put it here. 2 | 3 | ```angular2 4 | -SLM 5 | ---... 6 | ---predict 7 | -----... 8 | -----checkpoints 9 | -------AMFMN.tar 10 | -------RSITMD_LW_MCR.tar 11 | 12 | ``` 13 | -------------------------------------------------------------------------------- /predict/data/rsitmd_precomp/test_filename.txt: -------------------------------------------------------------------------------- 1 | boat_0.tif 2 | boat_0.tif 3 | boat_0.tif 4 | boat_0.tif 5 | boat_0.tif 6 | playground_1.tif 7 | playground_1.tif 8 | playground_1.tif 9 | playground_1.tif 10 | playground_1.tif 11 | airport_2.tif 12 | airport_2.tif 13 | airport_2.tif 14 | airport_2.tif 15 | airport_2.tif 16 | airport_3.tif 17 | airport_3.tif 18 | airport_3.tif 19 | airport_3.tif 20 | airport_3.tif 21 | airport_4.tif 22 | airport_4.tif 23 | airport_4.tif 24 | airport_4.tif 25 | airport_4.tif 26 | airport_5.tif 27 | airport_5.tif 28 | airport_5.tif 29 | airport_5.tif 30 | airport_5.tif 31 | airport_6.tif 32 | airport_6.tif 33 | airport_6.tif 34 | airport_6.tif 35 | airport_6.tif 36 | airport_7.tif 37 | airport_7.tif 38 | airport_7.tif 39 | airport_7.tif 40 | airport_7.tif 41 | airport_8.tif 42 | airport_8.tif 43 | airport_8.tif 44 | airport_8.tif 45 | airport_8.tif 46 | airport_9.tif 47 | airport_9.tif 48 | airport_9.tif 49 | airport_9.tif 50 | airport_9.tif 51 | airport_10.tif 52 | airport_10.tif 53 | airport_10.tif 54 | airport_10.tif 55 | airport_10.tif 56 | airport_11.tif 57 | airport_11.tif 58 | airport_11.tif 59 | airport_11.tif 60 | airport_11.tif 61 | airport_12.tif 62 | airport_12.tif 63 | airport_12.tif 64 | airport_12.tif 65 | airport_12.tif 66 | bareland_13.tif 67 | bareland_13.tif 68 | bareland_13.tif 69 | bareland_13.tif 70 | bareland_13.tif 71 | bareland_14.tif 72 | bareland_14.tif 73 | bareland_14.tif 74 | bareland_14.tif 75 | bareland_14.tif 76 | bareland_15.tif 77 | bareland_15.tif 78 | bareland_15.tif 79 | bareland_15.tif 80 | bareland_15.tif 81 | bareland_16.tif 82 | bareland_16.tif 83 | bareland_16.tif 84 | bareland_16.tif 85 | bareland_16.tif 86 | bareland_17.tif 87 | bareland_17.tif 88 | bareland_17.tif 89 | bareland_17.tif 90 | bareland_17.tif 91 | baseballfield_18.tif 92 | baseballfield_18.tif 93 | baseballfield_18.tif 94 | baseballfield_18.tif 95 | baseballfield_18.tif 96 | baseballfield_19.tif 97 | baseballfield_19.tif 98 | baseballfield_19.tif 99 | baseballfield_19.tif 100 | baseballfield_19.tif 101 | baseballfield_20.tif 102 | baseballfield_20.tif 103 | baseballfield_20.tif 104 | baseballfield_20.tif 105 | baseballfield_20.tif 106 | baseballfield_21.tif 107 | baseballfield_21.tif 108 | baseballfield_21.tif 109 | baseballfield_21.tif 110 | baseballfield_21.tif 111 | baseballfield_22.tif 112 | baseballfield_22.tif 113 | baseballfield_22.tif 114 | baseballfield_22.tif 115 | baseballfield_22.tif 116 | baseballfield_23.tif 117 | baseballfield_23.tif 118 | baseballfield_23.tif 119 | baseballfield_23.tif 120 | baseballfield_23.tif 121 | baseballfield_24.tif 122 | baseballfield_24.tif 123 | baseballfield_24.tif 124 | baseballfield_24.tif 125 | baseballfield_24.tif 126 | baseballfield_25.tif 127 | baseballfield_25.tif 128 | baseballfield_25.tif 129 | baseballfield_25.tif 130 | baseballfield_25.tif 131 | baseballfield_26.tif 132 | baseballfield_26.tif 133 | baseballfield_26.tif 134 | baseballfield_26.tif 135 | baseballfield_26.tif 136 | beach_27.tif 137 | beach_27.tif 138 | beach_27.tif 139 | beach_27.tif 140 | beach_27.tif 141 | beach_28.tif 142 | beach_28.tif 143 | beach_28.tif 144 | beach_28.tif 145 | beach_28.tif 146 | beach_29.tif 147 | beach_29.tif 148 | beach_29.tif 149 | beach_29.tif 150 | beach_29.tif 151 | beach_30.tif 152 | beach_30.tif 153 | beach_30.tif 154 | beach_30.tif 155 | beach_30.tif 156 | beach_31.tif 157 | beach_31.tif 158 | beach_31.tif 159 | beach_31.tif 160 | beach_31.tif 161 | beach_32.tif 162 | beach_32.tif 163 | beach_32.tif 164 | beach_32.tif 165 | beach_32.tif 166 | beach_33.tif 167 | beach_33.tif 168 | beach_33.tif 169 | beach_33.tif 170 | beach_33.tif 171 | beach_34.tif 172 | beach_34.tif 173 | beach_34.tif 174 | beach_34.tif 175 | beach_34.tif 176 | beach_35.tif 177 | beach_35.tif 178 | beach_35.tif 179 | beach_35.tif 180 | beach_35.tif 181 | beach_36.tif 182 | beach_36.tif 183 | beach_36.tif 184 | beach_36.tif 185 | beach_36.tif 186 | beach_37.tif 187 | beach_37.tif 188 | beach_37.tif 189 | beach_37.tif 190 | beach_37.tif 191 | beach_38.tif 192 | beach_38.tif 193 | beach_38.tif 194 | beach_38.tif 195 | beach_38.tif 196 | beach_39.tif 197 | beach_39.tif 198 | beach_39.tif 199 | beach_39.tif 200 | beach_39.tif 201 | beach_40.tif 202 | beach_40.tif 203 | beach_40.tif 204 | beach_40.tif 205 | beach_40.tif 206 | beach_41.tif 207 | beach_41.tif 208 | beach_41.tif 209 | beach_41.tif 210 | beach_41.tif 211 | bridge_42.tif 212 | bridge_42.tif 213 | bridge_42.tif 214 | bridge_42.tif 215 | bridge_42.tif 216 | bridge_43.tif 217 | bridge_43.tif 218 | bridge_43.tif 219 | bridge_43.tif 220 | bridge_43.tif 221 | bridge_44.tif 222 | bridge_44.tif 223 | bridge_44.tif 224 | bridge_44.tif 225 | bridge_44.tif 226 | bridge_45.tif 227 | bridge_45.tif 228 | bridge_45.tif 229 | bridge_45.tif 230 | bridge_45.tif 231 | bridge_46.tif 232 | bridge_46.tif 233 | bridge_46.tif 234 | bridge_46.tif 235 | bridge_46.tif 236 | bridge_47.tif 237 | bridge_47.tif 238 | bridge_47.tif 239 | bridge_47.tif 240 | bridge_47.tif 241 | bridge_48.tif 242 | bridge_48.tif 243 | bridge_48.tif 244 | bridge_48.tif 245 | bridge_48.tif 246 | bridge_49.tif 247 | bridge_49.tif 248 | bridge_49.tif 249 | bridge_49.tif 250 | bridge_49.tif 251 | bridge_50.tif 252 | bridge_50.tif 253 | bridge_50.tif 254 | bridge_50.tif 255 | bridge_50.tif 256 | bridge_51.tif 257 | bridge_51.tif 258 | bridge_51.tif 259 | bridge_51.tif 260 | bridge_51.tif 261 | bridge_52.tif 262 | bridge_52.tif 263 | bridge_52.tif 264 | bridge_52.tif 265 | bridge_52.tif 266 | bridge_53.tif 267 | bridge_53.tif 268 | bridge_53.tif 269 | bridge_53.tif 270 | bridge_53.tif 271 | bridge_54.tif 272 | bridge_54.tif 273 | bridge_54.tif 274 | bridge_54.tif 275 | bridge_54.tif 276 | bridge_55.tif 277 | bridge_55.tif 278 | bridge_55.tif 279 | bridge_55.tif 280 | bridge_55.tif 281 | bridge_56.tif 282 | bridge_56.tif 283 | bridge_56.tif 284 | bridge_56.tif 285 | bridge_56.tif 286 | center_57.tif 287 | center_57.tif 288 | center_57.tif 289 | center_57.tif 290 | center_57.tif 291 | center_58.tif 292 | center_58.tif 293 | center_58.tif 294 | center_58.tif 295 | center_58.tif 296 | center_59.tif 297 | center_59.tif 298 | center_59.tif 299 | center_59.tif 300 | center_59.tif 301 | center_60.tif 302 | center_60.tif 303 | center_60.tif 304 | center_60.tif 305 | center_60.tif 306 | center_61.tif 307 | center_61.tif 308 | center_61.tif 309 | center_61.tif 310 | center_61.tif 311 | center_62.tif 312 | center_62.tif 313 | center_62.tif 314 | center_62.tif 315 | center_62.tif 316 | center_63.tif 317 | center_63.tif 318 | center_63.tif 319 | center_63.tif 320 | center_63.tif 321 | center_64.tif 322 | center_64.tif 323 | center_64.tif 324 | center_64.tif 325 | center_64.tif 326 | center_65.tif 327 | center_65.tif 328 | center_65.tif 329 | center_65.tif 330 | center_65.tif 331 | center_66.tif 332 | center_66.tif 333 | center_66.tif 334 | center_66.tif 335 | center_66.tif 336 | center_67.tif 337 | center_67.tif 338 | center_67.tif 339 | center_67.tif 340 | center_67.tif 341 | center_68.tif 342 | center_68.tif 343 | center_68.tif 344 | center_68.tif 345 | center_68.tif 346 | center_69.tif 347 | center_69.tif 348 | center_69.tif 349 | center_69.tif 350 | center_69.tif 351 | center_70.tif 352 | center_70.tif 353 | center_70.tif 354 | center_70.tif 355 | center_70.tif 356 | church_71.tif 357 | church_71.tif 358 | church_71.tif 359 | church_71.tif 360 | church_71.tif 361 | church_72.tif 362 | church_72.tif 363 | church_72.tif 364 | church_72.tif 365 | church_72.tif 366 | church_73.tif 367 | church_73.tif 368 | church_73.tif 369 | church_73.tif 370 | church_73.tif 371 | church_74.tif 372 | church_74.tif 373 | church_74.tif 374 | church_74.tif 375 | church_74.tif 376 | church_75.tif 377 | church_75.tif 378 | church_75.tif 379 | church_75.tif 380 | church_75.tif 381 | church_76.tif 382 | church_76.tif 383 | church_76.tif 384 | church_76.tif 385 | church_76.tif 386 | church_77.tif 387 | church_77.tif 388 | church_77.tif 389 | church_77.tif 390 | church_77.tif 391 | church_78.tif 392 | church_78.tif 393 | church_78.tif 394 | church_78.tif 395 | church_78.tif 396 | church_79.tif 397 | church_79.tif 398 | church_79.tif 399 | church_79.tif 400 | church_79.tif 401 | church_80.tif 402 | church_80.tif 403 | church_80.tif 404 | church_80.tif 405 | church_80.tif 406 | church_81.tif 407 | church_81.tif 408 | church_81.tif 409 | church_81.tif 410 | church_81.tif 411 | church_82.tif 412 | church_82.tif 413 | church_82.tif 414 | church_82.tif 415 | church_82.tif 416 | church_83.tif 417 | church_83.tif 418 | church_83.tif 419 | church_83.tif 420 | church_83.tif 421 | commercial_84.tif 422 | commercial_84.tif 423 | commercial_84.tif 424 | commercial_84.tif 425 | commercial_84.tif 426 | commercial_85.tif 427 | commercial_85.tif 428 | commercial_85.tif 429 | commercial_85.tif 430 | commercial_85.tif 431 | commercial_86.tif 432 | commercial_86.tif 433 | commercial_86.tif 434 | commercial_86.tif 435 | commercial_86.tif 436 | commercial_87.tif 437 | commercial_87.tif 438 | commercial_87.tif 439 | commercial_87.tif 440 | commercial_87.tif 441 | commercial_88.tif 442 | commercial_88.tif 443 | commercial_88.tif 444 | commercial_88.tif 445 | commercial_88.tif 446 | commercial_89.tif 447 | commercial_89.tif 448 | commercial_89.tif 449 | commercial_89.tif 450 | commercial_89.tif 451 | commercial_90.tif 452 | commercial_90.tif 453 | commercial_90.tif 454 | commercial_90.tif 455 | commercial_90.tif 456 | commercial_91.tif 457 | commercial_91.tif 458 | commercial_91.tif 459 | commercial_91.tif 460 | commercial_91.tif 461 | commercial_92.tif 462 | commercial_92.tif 463 | commercial_92.tif 464 | commercial_92.tif 465 | commercial_92.tif 466 | commercial_93.tif 467 | commercial_93.tif 468 | commercial_93.tif 469 | commercial_93.tif 470 | commercial_93.tif 471 | commercial_94.tif 472 | commercial_94.tif 473 | commercial_94.tif 474 | commercial_94.tif 475 | commercial_94.tif 476 | commercial_95.tif 477 | commercial_95.tif 478 | commercial_95.tif 479 | commercial_95.tif 480 | commercial_95.tif 481 | commercial_96.tif 482 | commercial_96.tif 483 | commercial_96.tif 484 | commercial_96.tif 485 | commercial_96.tif 486 | commercial_97.tif 487 | commercial_97.tif 488 | commercial_97.tif 489 | commercial_97.tif 490 | commercial_97.tif 491 | denseresidential_98.tif 492 | denseresidential_98.tif 493 | denseresidential_98.tif 494 | denseresidential_98.tif 495 | denseresidential_98.tif 496 | denseresidential_99.tif 497 | denseresidential_99.tif 498 | denseresidential_99.tif 499 | denseresidential_99.tif 500 | denseresidential_99.tif 501 | denseresidential_100.tif 502 | denseresidential_100.tif 503 | denseresidential_100.tif 504 | denseresidential_100.tif 505 | denseresidential_100.tif 506 | denseresidential_101.tif 507 | denseresidential_101.tif 508 | denseresidential_101.tif 509 | denseresidential_101.tif 510 | denseresidential_101.tif 511 | denseresidential_102.tif 512 | denseresidential_102.tif 513 | denseresidential_102.tif 514 | denseresidential_102.tif 515 | denseresidential_102.tif 516 | denseresidential_103.tif 517 | denseresidential_103.tif 518 | denseresidential_103.tif 519 | denseresidential_103.tif 520 | denseresidential_103.tif 521 | denseresidential_104.tif 522 | denseresidential_104.tif 523 | denseresidential_104.tif 524 | denseresidential_104.tif 525 | denseresidential_104.tif 526 | denseresidential_105.tif 527 | denseresidential_105.tif 528 | denseresidential_105.tif 529 | denseresidential_105.tif 530 | denseresidential_105.tif 531 | denseresidential_106.tif 532 | denseresidential_106.tif 533 | denseresidential_106.tif 534 | denseresidential_106.tif 535 | denseresidential_106.tif 536 | denseresidential_107.tif 537 | denseresidential_107.tif 538 | denseresidential_107.tif 539 | denseresidential_107.tif 540 | denseresidential_107.tif 541 | denseresidential_108.tif 542 | denseresidential_108.tif 543 | denseresidential_108.tif 544 | denseresidential_108.tif 545 | denseresidential_108.tif 546 | denseresidential_109.tif 547 | denseresidential_109.tif 548 | denseresidential_109.tif 549 | denseresidential_109.tif 550 | denseresidential_109.tif 551 | denseresidential_110.tif 552 | denseresidential_110.tif 553 | denseresidential_110.tif 554 | denseresidential_110.tif 555 | denseresidential_110.tif 556 | denseresidential_111.tif 557 | denseresidential_111.tif 558 | denseresidential_111.tif 559 | denseresidential_111.tif 560 | denseresidential_111.tif 561 | desert_112.tif 562 | desert_112.tif 563 | desert_112.tif 564 | desert_112.tif 565 | desert_112.tif 566 | desert_113.tif 567 | desert_113.tif 568 | desert_113.tif 569 | desert_113.tif 570 | desert_113.tif 571 | desert_114.tif 572 | desert_114.tif 573 | desert_114.tif 574 | desert_114.tif 575 | desert_114.tif 576 | desert_115.tif 577 | desert_115.tif 578 | desert_115.tif 579 | desert_115.tif 580 | desert_115.tif 581 | desert_116.tif 582 | desert_116.tif 583 | desert_116.tif 584 | desert_116.tif 585 | desert_116.tif 586 | desert_117.tif 587 | desert_117.tif 588 | desert_117.tif 589 | desert_117.tif 590 | desert_117.tif 591 | desert_118.tif 592 | desert_118.tif 593 | desert_118.tif 594 | desert_118.tif 595 | desert_118.tif 596 | desert_119.tif 597 | desert_119.tif 598 | desert_119.tif 599 | desert_119.tif 600 | desert_119.tif 601 | desert_120.tif 602 | desert_120.tif 603 | desert_120.tif 604 | desert_120.tif 605 | desert_120.tif 606 | desert_121.tif 607 | desert_121.tif 608 | desert_121.tif 609 | desert_121.tif 610 | desert_121.tif 611 | desert_122.tif 612 | desert_122.tif 613 | desert_122.tif 614 | desert_122.tif 615 | desert_122.tif 616 | desert_123.tif 617 | desert_123.tif 618 | desert_123.tif 619 | desert_123.tif 620 | desert_123.tif 621 | farmland_124.tif 622 | farmland_124.tif 623 | farmland_124.tif 624 | farmland_124.tif 625 | farmland_124.tif 626 | farmland_125.tif 627 | farmland_125.tif 628 | farmland_125.tif 629 | farmland_125.tif 630 | farmland_125.tif 631 | farmland_126.tif 632 | farmland_126.tif 633 | farmland_126.tif 634 | farmland_126.tif 635 | farmland_126.tif 636 | farmland_127.tif 637 | farmland_127.tif 638 | farmland_127.tif 639 | farmland_127.tif 640 | farmland_127.tif 641 | farmland_128.tif 642 | farmland_128.tif 643 | farmland_128.tif 644 | farmland_128.tif 645 | farmland_128.tif 646 | farmland_129.tif 647 | farmland_129.tif 648 | farmland_129.tif 649 | farmland_129.tif 650 | farmland_129.tif 651 | farmland_130.tif 652 | farmland_130.tif 653 | farmland_130.tif 654 | farmland_130.tif 655 | farmland_130.tif 656 | farmland_131.tif 657 | farmland_131.tif 658 | farmland_131.tif 659 | farmland_131.tif 660 | farmland_131.tif 661 | farmland_132.tif 662 | farmland_132.tif 663 | farmland_132.tif 664 | farmland_132.tif 665 | farmland_132.tif 666 | farmland_133.tif 667 | farmland_133.tif 668 | farmland_133.tif 669 | farmland_133.tif 670 | farmland_133.tif 671 | farmland_134.tif 672 | farmland_134.tif 673 | farmland_134.tif 674 | farmland_134.tif 675 | farmland_134.tif 676 | farmland_135.tif 677 | farmland_135.tif 678 | farmland_135.tif 679 | farmland_135.tif 680 | farmland_135.tif 681 | farmland_136.tif 682 | farmland_136.tif 683 | farmland_136.tif 684 | farmland_136.tif 685 | farmland_136.tif 686 | farmland_137.tif 687 | farmland_137.tif 688 | farmland_137.tif 689 | farmland_137.tif 690 | farmland_137.tif 691 | farmland_138.tif 692 | farmland_138.tif 693 | farmland_138.tif 694 | farmland_138.tif 695 | farmland_138.tif 696 | farmland_139.tif 697 | farmland_139.tif 698 | farmland_139.tif 699 | farmland_139.tif 700 | farmland_139.tif 701 | farmland_140.tif 702 | farmland_140.tif 703 | farmland_140.tif 704 | farmland_140.tif 705 | farmland_140.tif 706 | forest_141.tif 707 | forest_141.tif 708 | forest_141.tif 709 | forest_141.tif 710 | forest_141.tif 711 | forest_142.tif 712 | forest_142.tif 713 | forest_142.tif 714 | forest_142.tif 715 | forest_142.tif 716 | forest_143.tif 717 | forest_143.tif 718 | forest_143.tif 719 | forest_143.tif 720 | forest_143.tif 721 | forest_144.tif 722 | forest_144.tif 723 | forest_144.tif 724 | forest_144.tif 725 | forest_144.tif 726 | forest_145.tif 727 | forest_145.tif 728 | forest_145.tif 729 | forest_145.tif 730 | forest_145.tif 731 | forest_146.tif 732 | forest_146.tif 733 | forest_146.tif 734 | forest_146.tif 735 | forest_146.tif 736 | forest_147.tif 737 | forest_147.tif 738 | forest_147.tif 739 | forest_147.tif 740 | forest_147.tif 741 | forest_148.tif 742 | forest_148.tif 743 | forest_148.tif 744 | forest_148.tif 745 | forest_148.tif 746 | forest_149.tif 747 | forest_149.tif 748 | forest_149.tif 749 | forest_149.tif 750 | forest_149.tif 751 | forest_150.tif 752 | forest_150.tif 753 | forest_150.tif 754 | forest_150.tif 755 | forest_150.tif 756 | forest_151.tif 757 | forest_151.tif 758 | forest_151.tif 759 | forest_151.tif 760 | forest_151.tif 761 | industrial_152.tif 762 | industrial_152.tif 763 | industrial_152.tif 764 | industrial_152.tif 765 | industrial_152.tif 766 | industrial_153.tif 767 | industrial_153.tif 768 | industrial_153.tif 769 | industrial_153.tif 770 | industrial_153.tif 771 | industrial_154.tif 772 | industrial_154.tif 773 | industrial_154.tif 774 | industrial_154.tif 775 | industrial_154.tif 776 | industrial_155.tif 777 | industrial_155.tif 778 | industrial_155.tif 779 | industrial_155.tif 780 | industrial_155.tif 781 | industrial_156.tif 782 | industrial_156.tif 783 | industrial_156.tif 784 | industrial_156.tif 785 | industrial_156.tif 786 | industrial_157.tif 787 | industrial_157.tif 788 | industrial_157.tif 789 | industrial_157.tif 790 | industrial_157.tif 791 | industrial_158.tif 792 | industrial_158.tif 793 | industrial_158.tif 794 | industrial_158.tif 795 | industrial_158.tif 796 | industrial_159.tif 797 | industrial_159.tif 798 | industrial_159.tif 799 | industrial_159.tif 800 | industrial_159.tif 801 | industrial_160.tif 802 | industrial_160.tif 803 | industrial_160.tif 804 | industrial_160.tif 805 | industrial_160.tif 806 | industrial_161.tif 807 | industrial_161.tif 808 | industrial_161.tif 809 | industrial_161.tif 810 | industrial_161.tif 811 | industrial_162.tif 812 | industrial_162.tif 813 | industrial_162.tif 814 | industrial_162.tif 815 | industrial_162.tif 816 | industrial_163.tif 817 | industrial_163.tif 818 | industrial_163.tif 819 | industrial_163.tif 820 | industrial_163.tif 821 | industrial_164.tif 822 | industrial_164.tif 823 | industrial_164.tif 824 | industrial_164.tif 825 | industrial_164.tif 826 | industrial_165.tif 827 | industrial_165.tif 828 | industrial_165.tif 829 | industrial_165.tif 830 | industrial_165.tif 831 | industrial_166.tif 832 | industrial_166.tif 833 | industrial_166.tif 834 | industrial_166.tif 835 | industrial_166.tif 836 | industrial_167.tif 837 | industrial_167.tif 838 | industrial_167.tif 839 | industrial_167.tif 840 | industrial_167.tif 841 | industrial_168.tif 842 | industrial_168.tif 843 | industrial_168.tif 844 | industrial_168.tif 845 | industrial_168.tif 846 | industrial_169.tif 847 | industrial_169.tif 848 | industrial_169.tif 849 | industrial_169.tif 850 | industrial_169.tif 851 | industrial_170.tif 852 | industrial_170.tif 853 | industrial_170.tif 854 | industrial_170.tif 855 | industrial_170.tif 856 | industrial_171.tif 857 | industrial_171.tif 858 | industrial_171.tif 859 | industrial_171.tif 860 | industrial_171.tif 861 | meadow_172.tif 862 | meadow_172.tif 863 | meadow_172.tif 864 | meadow_172.tif 865 | meadow_172.tif 866 | meadow_173.tif 867 | meadow_173.tif 868 | meadow_173.tif 869 | meadow_173.tif 870 | meadow_173.tif 871 | meadow_174.tif 872 | meadow_174.tif 873 | meadow_174.tif 874 | meadow_174.tif 875 | meadow_174.tif 876 | meadow_175.tif 877 | meadow_175.tif 878 | meadow_175.tif 879 | meadow_175.tif 880 | meadow_175.tif 881 | meadow_176.tif 882 | meadow_176.tif 883 | meadow_176.tif 884 | meadow_176.tif 885 | meadow_176.tif 886 | meadow_177.tif 887 | meadow_177.tif 888 | meadow_177.tif 889 | meadow_177.tif 890 | meadow_177.tif 891 | meadow_178.tif 892 | meadow_178.tif 893 | meadow_178.tif 894 | meadow_178.tif 895 | meadow_178.tif 896 | meadow_179.tif 897 | meadow_179.tif 898 | meadow_179.tif 899 | meadow_179.tif 900 | meadow_179.tif 901 | meadow_180.tif 902 | meadow_180.tif 903 | meadow_180.tif 904 | meadow_180.tif 905 | meadow_180.tif 906 | meadow_181.tif 907 | meadow_181.tif 908 | meadow_181.tif 909 | meadow_181.tif 910 | meadow_181.tif 911 | meadow_182.tif 912 | meadow_182.tif 913 | meadow_182.tif 914 | meadow_182.tif 915 | meadow_182.tif 916 | meadow_183.tif 917 | meadow_183.tif 918 | meadow_183.tif 919 | meadow_183.tif 920 | meadow_183.tif 921 | mediumresidential_184.tif 922 | mediumresidential_184.tif 923 | mediumresidential_184.tif 924 | mediumresidential_184.tif 925 | mediumresidential_184.tif 926 | mediumresidential_185.tif 927 | mediumresidential_185.tif 928 | mediumresidential_185.tif 929 | mediumresidential_185.tif 930 | mediumresidential_185.tif 931 | mediumresidential_186.tif 932 | mediumresidential_186.tif 933 | mediumresidential_186.tif 934 | mediumresidential_186.tif 935 | mediumresidential_186.tif 936 | mediumresidential_187.tif 937 | mediumresidential_187.tif 938 | mediumresidential_187.tif 939 | mediumresidential_187.tif 940 | mediumresidential_187.tif 941 | mediumresidential_188.tif 942 | mediumresidential_188.tif 943 | mediumresidential_188.tif 944 | mediumresidential_188.tif 945 | mediumresidential_188.tif 946 | mediumresidential_189.tif 947 | mediumresidential_189.tif 948 | mediumresidential_189.tif 949 | mediumresidential_189.tif 950 | mediumresidential_189.tif 951 | mediumresidential_190.tif 952 | mediumresidential_190.tif 953 | mediumresidential_190.tif 954 | mediumresidential_190.tif 955 | mediumresidential_190.tif 956 | mediumresidential_191.tif 957 | mediumresidential_191.tif 958 | mediumresidential_191.tif 959 | mediumresidential_191.tif 960 | mediumresidential_191.tif 961 | mediumresidential_192.tif 962 | mediumresidential_192.tif 963 | mediumresidential_192.tif 964 | mediumresidential_192.tif 965 | mediumresidential_192.tif 966 | mediumresidential_193.tif 967 | mediumresidential_193.tif 968 | mediumresidential_193.tif 969 | mediumresidential_193.tif 970 | mediumresidential_193.tif 971 | mediumresidential_194.tif 972 | mediumresidential_194.tif 973 | mediumresidential_194.tif 974 | mediumresidential_194.tif 975 | mediumresidential_194.tif 976 | mediumresidential_195.tif 977 | mediumresidential_195.tif 978 | mediumresidential_195.tif 979 | mediumresidential_195.tif 980 | mediumresidential_195.tif 981 | mediumresidential_196.tif 982 | mediumresidential_196.tif 983 | mediumresidential_196.tif 984 | mediumresidential_196.tif 985 | mediumresidential_196.tif 986 | mountain_197.tif 987 | mountain_197.tif 988 | mountain_197.tif 989 | mountain_197.tif 990 | mountain_197.tif 991 | mountain_198.tif 992 | mountain_198.tif 993 | mountain_198.tif 994 | mountain_198.tif 995 | mountain_198.tif 996 | mountain_199.tif 997 | mountain_199.tif 998 | mountain_199.tif 999 | mountain_199.tif 1000 | mountain_199.tif 1001 | mountain_200.tif 1002 | mountain_200.tif 1003 | mountain_200.tif 1004 | mountain_200.tif 1005 | mountain_200.tif 1006 | mountain_201.tif 1007 | mountain_201.tif 1008 | mountain_201.tif 1009 | mountain_201.tif 1010 | mountain_201.tif 1011 | mountain_202.tif 1012 | mountain_202.tif 1013 | mountain_202.tif 1014 | mountain_202.tif 1015 | mountain_202.tif 1016 | mountain_203.tif 1017 | mountain_203.tif 1018 | mountain_203.tif 1019 | mountain_203.tif 1020 | mountain_203.tif 1021 | mountain_204.tif 1022 | mountain_204.tif 1023 | mountain_204.tif 1024 | mountain_204.tif 1025 | mountain_204.tif 1026 | mountain_205.tif 1027 | mountain_205.tif 1028 | mountain_205.tif 1029 | mountain_205.tif 1030 | mountain_205.tif 1031 | mountain_206.tif 1032 | mountain_206.tif 1033 | mountain_206.tif 1034 | mountain_206.tif 1035 | mountain_206.tif 1036 | park_207.tif 1037 | park_207.tif 1038 | park_207.tif 1039 | park_207.tif 1040 | park_207.tif 1041 | park_208.tif 1042 | park_208.tif 1043 | park_208.tif 1044 | park_208.tif 1045 | park_208.tif 1046 | park_209.tif 1047 | park_209.tif 1048 | park_209.tif 1049 | park_209.tif 1050 | park_209.tif 1051 | park_210.tif 1052 | park_210.tif 1053 | park_210.tif 1054 | park_210.tif 1055 | park_210.tif 1056 | park_211.tif 1057 | park_211.tif 1058 | park_211.tif 1059 | park_211.tif 1060 | park_211.tif 1061 | park_212.tif 1062 | park_212.tif 1063 | park_212.tif 1064 | park_212.tif 1065 | park_212.tif 1066 | park_213.tif 1067 | park_213.tif 1068 | park_213.tif 1069 | park_213.tif 1070 | park_213.tif 1071 | park_214.tif 1072 | park_214.tif 1073 | park_214.tif 1074 | park_214.tif 1075 | park_214.tif 1076 | park_215.tif 1077 | park_215.tif 1078 | park_215.tif 1079 | park_215.tif 1080 | park_215.tif 1081 | park_216.tif 1082 | park_216.tif 1083 | park_216.tif 1084 | park_216.tif 1085 | park_216.tif 1086 | park_217.tif 1087 | park_217.tif 1088 | park_217.tif 1089 | park_217.tif 1090 | park_217.tif 1091 | park_218.tif 1092 | park_218.tif 1093 | park_218.tif 1094 | park_218.tif 1095 | park_218.tif 1096 | park_219.tif 1097 | park_219.tif 1098 | park_219.tif 1099 | park_219.tif 1100 | park_219.tif 1101 | park_220.tif 1102 | park_220.tif 1103 | park_220.tif 1104 | park_220.tif 1105 | park_220.tif 1106 | park_221.tif 1107 | park_221.tif 1108 | park_221.tif 1109 | park_221.tif 1110 | park_221.tif 1111 | park_222.tif 1112 | park_222.tif 1113 | park_222.tif 1114 | park_222.tif 1115 | park_222.tif 1116 | parking_223.tif 1117 | parking_223.tif 1118 | parking_223.tif 1119 | parking_223.tif 1120 | parking_223.tif 1121 | parking_224.tif 1122 | parking_224.tif 1123 | parking_224.tif 1124 | parking_224.tif 1125 | parking_224.tif 1126 | parking_225.tif 1127 | parking_225.tif 1128 | parking_225.tif 1129 | parking_225.tif 1130 | parking_225.tif 1131 | parking_226.tif 1132 | parking_226.tif 1133 | parking_226.tif 1134 | parking_226.tif 1135 | parking_226.tif 1136 | parking_227.tif 1137 | parking_227.tif 1138 | parking_227.tif 1139 | parking_227.tif 1140 | parking_227.tif 1141 | parking_228.tif 1142 | parking_228.tif 1143 | parking_228.tif 1144 | parking_228.tif 1145 | parking_228.tif 1146 | parking_229.tif 1147 | parking_229.tif 1148 | parking_229.tif 1149 | parking_229.tif 1150 | parking_229.tif 1151 | parking_230.tif 1152 | parking_230.tif 1153 | parking_230.tif 1154 | parking_230.tif 1155 | parking_230.tif 1156 | parking_231.tif 1157 | parking_231.tif 1158 | parking_231.tif 1159 | parking_231.tif 1160 | parking_231.tif 1161 | parking_232.tif 1162 | parking_232.tif 1163 | parking_232.tif 1164 | parking_232.tif 1165 | parking_232.tif 1166 | parking_233.tif 1167 | parking_233.tif 1168 | parking_233.tif 1169 | parking_233.tif 1170 | parking_233.tif 1171 | parking_234.tif 1172 | parking_234.tif 1173 | parking_234.tif 1174 | parking_234.tif 1175 | parking_234.tif 1176 | parking_235.tif 1177 | parking_235.tif 1178 | parking_235.tif 1179 | parking_235.tif 1180 | parking_235.tif 1181 | parking_236.tif 1182 | parking_236.tif 1183 | parking_236.tif 1184 | parking_236.tif 1185 | parking_236.tif 1186 | playground_237.tif 1187 | playground_237.tif 1188 | playground_237.tif 1189 | playground_237.tif 1190 | playground_237.tif 1191 | playground_238.tif 1192 | playground_238.tif 1193 | playground_238.tif 1194 | playground_238.tif 1195 | playground_238.tif 1196 | playground_239.tif 1197 | playground_239.tif 1198 | playground_239.tif 1199 | playground_239.tif 1200 | playground_239.tif 1201 | playground_240.tif 1202 | playground_240.tif 1203 | playground_240.tif 1204 | playground_240.tif 1205 | playground_240.tif 1206 | playground_241.tif 1207 | playground_241.tif 1208 | playground_241.tif 1209 | playground_241.tif 1210 | playground_241.tif 1211 | playground_242.tif 1212 | playground_242.tif 1213 | playground_242.tif 1214 | playground_242.tif 1215 | playground_242.tif 1216 | playground_243.tif 1217 | playground_243.tif 1218 | playground_243.tif 1219 | playground_243.tif 1220 | playground_243.tif 1221 | playground_244.tif 1222 | playground_244.tif 1223 | playground_244.tif 1224 | playground_244.tif 1225 | playground_244.tif 1226 | playground_245.tif 1227 | playground_245.tif 1228 | playground_245.tif 1229 | playground_245.tif 1230 | playground_245.tif 1231 | playground_246.tif 1232 | playground_246.tif 1233 | playground_246.tif 1234 | playground_246.tif 1235 | playground_246.tif 1236 | playground_247.tif 1237 | playground_247.tif 1238 | playground_247.tif 1239 | playground_247.tif 1240 | playground_247.tif 1241 | playground_248.tif 1242 | playground_248.tif 1243 | playground_248.tif 1244 | playground_248.tif 1245 | playground_248.tif 1246 | playground_249.tif 1247 | playground_249.tif 1248 | playground_249.tif 1249 | playground_249.tif 1250 | playground_249.tif 1251 | playground_250.tif 1252 | playground_250.tif 1253 | playground_250.tif 1254 | playground_250.tif 1255 | playground_250.tif 1256 | playground_251.tif 1257 | playground_251.tif 1258 | playground_251.tif 1259 | playground_251.tif 1260 | playground_251.tif 1261 | playground_252.tif 1262 | playground_252.tif 1263 | playground_252.tif 1264 | playground_252.tif 1265 | playground_252.tif 1266 | pond_253.tif 1267 | pond_253.tif 1268 | pond_253.tif 1269 | pond_253.tif 1270 | pond_253.tif 1271 | pond_254.tif 1272 | pond_254.tif 1273 | pond_254.tif 1274 | pond_254.tif 1275 | pond_254.tif 1276 | pond_255.tif 1277 | pond_255.tif 1278 | pond_255.tif 1279 | pond_255.tif 1280 | pond_255.tif 1281 | pond_256.tif 1282 | pond_256.tif 1283 | pond_256.tif 1284 | pond_256.tif 1285 | pond_256.tif 1286 | pond_257.tif 1287 | pond_257.tif 1288 | pond_257.tif 1289 | pond_257.tif 1290 | pond_257.tif 1291 | pond_258.tif 1292 | pond_258.tif 1293 | pond_258.tif 1294 | pond_258.tif 1295 | pond_258.tif 1296 | pond_259.tif 1297 | pond_259.tif 1298 | pond_259.tif 1299 | pond_259.tif 1300 | pond_259.tif 1301 | pond_260.tif 1302 | pond_260.tif 1303 | pond_260.tif 1304 | pond_260.tif 1305 | pond_260.tif 1306 | pond_261.tif 1307 | pond_261.tif 1308 | pond_261.tif 1309 | pond_261.tif 1310 | pond_261.tif 1311 | pond_262.tif 1312 | pond_262.tif 1313 | pond_262.tif 1314 | pond_262.tif 1315 | pond_262.tif 1316 | pond_263.tif 1317 | pond_263.tif 1318 | pond_263.tif 1319 | pond_263.tif 1320 | pond_263.tif 1321 | pond_264.tif 1322 | pond_264.tif 1323 | pond_264.tif 1324 | pond_264.tif 1325 | pond_264.tif 1326 | pond_265.tif 1327 | pond_265.tif 1328 | pond_265.tif 1329 | pond_265.tif 1330 | pond_265.tif 1331 | pond_266.tif 1332 | pond_266.tif 1333 | pond_266.tif 1334 | pond_266.tif 1335 | pond_266.tif 1336 | pond_267.tif 1337 | pond_267.tif 1338 | pond_267.tif 1339 | pond_267.tif 1340 | pond_267.tif 1341 | pond_268.tif 1342 | pond_268.tif 1343 | pond_268.tif 1344 | pond_268.tif 1345 | pond_268.tif 1346 | pond_269.tif 1347 | pond_269.tif 1348 | pond_269.tif 1349 | pond_269.tif 1350 | pond_269.tif 1351 | pond_270.tif 1352 | pond_270.tif 1353 | pond_270.tif 1354 | pond_270.tif 1355 | pond_270.tif 1356 | pond_271.tif 1357 | pond_271.tif 1358 | pond_271.tif 1359 | pond_271.tif 1360 | pond_271.tif 1361 | pond_272.tif 1362 | pond_272.tif 1363 | pond_272.tif 1364 | pond_272.tif 1365 | pond_272.tif 1366 | pond_273.tif 1367 | pond_273.tif 1368 | pond_273.tif 1369 | pond_273.tif 1370 | pond_273.tif 1371 | pond_274.tif 1372 | pond_274.tif 1373 | pond_274.tif 1374 | pond_274.tif 1375 | pond_274.tif 1376 | port_275.tif 1377 | port_275.tif 1378 | port_275.tif 1379 | port_275.tif 1380 | port_275.tif 1381 | port_276.tif 1382 | port_276.tif 1383 | port_276.tif 1384 | port_276.tif 1385 | port_276.tif 1386 | port_277.tif 1387 | port_277.tif 1388 | port_277.tif 1389 | port_277.tif 1390 | port_277.tif 1391 | port_278.tif 1392 | port_278.tif 1393 | port_278.tif 1394 | port_278.tif 1395 | port_278.tif 1396 | port_279.tif 1397 | port_279.tif 1398 | port_279.tif 1399 | port_279.tif 1400 | port_279.tif 1401 | port_280.tif 1402 | port_280.tif 1403 | port_280.tif 1404 | port_280.tif 1405 | port_280.tif 1406 | port_281.tif 1407 | port_281.tif 1408 | port_281.tif 1409 | port_281.tif 1410 | port_281.tif 1411 | port_282.tif 1412 | port_282.tif 1413 | port_282.tif 1414 | port_282.tif 1415 | port_282.tif 1416 | port_283.tif 1417 | port_283.tif 1418 | port_283.tif 1419 | port_283.tif 1420 | port_283.tif 1421 | port_284.tif 1422 | port_284.tif 1423 | port_284.tif 1424 | port_284.tif 1425 | port_284.tif 1426 | port_285.tif 1427 | port_285.tif 1428 | port_285.tif 1429 | port_285.tif 1430 | port_285.tif 1431 | port_286.tif 1432 | port_286.tif 1433 | port_286.tif 1434 | port_286.tif 1435 | port_286.tif 1436 | port_287.tif 1437 | port_287.tif 1438 | port_287.tif 1439 | port_287.tif 1440 | port_287.tif 1441 | port_288.tif 1442 | port_288.tif 1443 | port_288.tif 1444 | port_288.tif 1445 | port_288.tif 1446 | port_289.tif 1447 | port_289.tif 1448 | port_289.tif 1449 | port_289.tif 1450 | port_289.tif 1451 | port_290.tif 1452 | port_290.tif 1453 | port_290.tif 1454 | port_290.tif 1455 | port_290.tif 1456 | port_291.tif 1457 | port_291.tif 1458 | port_291.tif 1459 | port_291.tif 1460 | port_291.tif 1461 | railwaystation_292.tif 1462 | railwaystation_292.tif 1463 | railwaystation_292.tif 1464 | railwaystation_292.tif 1465 | railwaystation_292.tif 1466 | railwaystation_293.tif 1467 | railwaystation_293.tif 1468 | railwaystation_293.tif 1469 | railwaystation_293.tif 1470 | railwaystation_293.tif 1471 | railwaystation_294.tif 1472 | railwaystation_294.tif 1473 | railwaystation_294.tif 1474 | railwaystation_294.tif 1475 | railwaystation_294.tif 1476 | railwaystation_295.tif 1477 | railwaystation_295.tif 1478 | railwaystation_295.tif 1479 | railwaystation_295.tif 1480 | railwaystation_295.tif 1481 | railwaystation_296.tif 1482 | railwaystation_296.tif 1483 | railwaystation_296.tif 1484 | railwaystation_296.tif 1485 | railwaystation_296.tif 1486 | railwaystation_297.tif 1487 | railwaystation_297.tif 1488 | railwaystation_297.tif 1489 | railwaystation_297.tif 1490 | railwaystation_297.tif 1491 | railwaystation_298.tif 1492 | railwaystation_298.tif 1493 | railwaystation_298.tif 1494 | railwaystation_298.tif 1495 | railwaystation_298.tif 1496 | railwaystation_299.tif 1497 | railwaystation_299.tif 1498 | railwaystation_299.tif 1499 | railwaystation_299.tif 1500 | railwaystation_299.tif 1501 | railwaystation_300.tif 1502 | railwaystation_300.tif 1503 | railwaystation_300.tif 1504 | railwaystation_300.tif 1505 | railwaystation_300.tif 1506 | railwaystation_301.tif 1507 | railwaystation_301.tif 1508 | railwaystation_301.tif 1509 | railwaystation_301.tif 1510 | railwaystation_301.tif 1511 | railwaystation_302.tif 1512 | railwaystation_302.tif 1513 | railwaystation_302.tif 1514 | railwaystation_302.tif 1515 | railwaystation_302.tif 1516 | resort_303.tif 1517 | resort_303.tif 1518 | resort_303.tif 1519 | resort_303.tif 1520 | resort_303.tif 1521 | resort_304.tif 1522 | resort_304.tif 1523 | resort_304.tif 1524 | resort_304.tif 1525 | resort_304.tif 1526 | resort_305.tif 1527 | resort_305.tif 1528 | resort_305.tif 1529 | resort_305.tif 1530 | resort_305.tif 1531 | resort_306.tif 1532 | resort_306.tif 1533 | resort_306.tif 1534 | resort_306.tif 1535 | resort_306.tif 1536 | resort_307.tif 1537 | resort_307.tif 1538 | resort_307.tif 1539 | resort_307.tif 1540 | resort_307.tif 1541 | resort_308.tif 1542 | resort_308.tif 1543 | resort_308.tif 1544 | resort_308.tif 1545 | resort_308.tif 1546 | resort_309.tif 1547 | resort_309.tif 1548 | resort_309.tif 1549 | resort_309.tif 1550 | resort_309.tif 1551 | resort_310.tif 1552 | resort_310.tif 1553 | resort_310.tif 1554 | resort_310.tif 1555 | resort_310.tif 1556 | resort_311.tif 1557 | resort_311.tif 1558 | resort_311.tif 1559 | resort_311.tif 1560 | resort_311.tif 1561 | resort_312.tif 1562 | resort_312.tif 1563 | resort_312.tif 1564 | resort_312.tif 1565 | resort_312.tif 1566 | resort_313.tif 1567 | resort_313.tif 1568 | resort_313.tif 1569 | resort_313.tif 1570 | resort_313.tif 1571 | resort_314.tif 1572 | resort_314.tif 1573 | resort_314.tif 1574 | resort_314.tif 1575 | resort_314.tif 1576 | resort_315.tif 1577 | resort_315.tif 1578 | resort_315.tif 1579 | resort_315.tif 1580 | resort_315.tif 1581 | resort_316.tif 1582 | resort_316.tif 1583 | resort_316.tif 1584 | resort_316.tif 1585 | resort_316.tif 1586 | resort_317.tif 1587 | resort_317.tif 1588 | resort_317.tif 1589 | resort_317.tif 1590 | resort_317.tif 1591 | resort_318.tif 1592 | resort_318.tif 1593 | resort_318.tif 1594 | resort_318.tif 1595 | resort_318.tif 1596 | river_319.tif 1597 | river_319.tif 1598 | river_319.tif 1599 | river_319.tif 1600 | river_319.tif 1601 | river_320.tif 1602 | river_320.tif 1603 | river_320.tif 1604 | river_320.tif 1605 | river_320.tif 1606 | river_321.tif 1607 | river_321.tif 1608 | river_321.tif 1609 | river_321.tif 1610 | river_321.tif 1611 | river_322.tif 1612 | river_322.tif 1613 | river_322.tif 1614 | river_322.tif 1615 | river_322.tif 1616 | river_323.tif 1617 | river_323.tif 1618 | river_323.tif 1619 | river_323.tif 1620 | river_323.tif 1621 | river_324.tif 1622 | river_324.tif 1623 | river_324.tif 1624 | river_324.tif 1625 | river_324.tif 1626 | river_325.tif 1627 | river_325.tif 1628 | river_325.tif 1629 | river_325.tif 1630 | river_325.tif 1631 | river_326.tif 1632 | river_326.tif 1633 | river_326.tif 1634 | river_326.tif 1635 | river_326.tif 1636 | river_327.tif 1637 | river_327.tif 1638 | river_327.tif 1639 | river_327.tif 1640 | river_327.tif 1641 | river_328.tif 1642 | river_328.tif 1643 | river_328.tif 1644 | river_328.tif 1645 | river_328.tif 1646 | river_329.tif 1647 | river_329.tif 1648 | river_329.tif 1649 | river_329.tif 1650 | river_329.tif 1651 | river_330.tif 1652 | river_330.tif 1653 | river_330.tif 1654 | river_330.tif 1655 | river_330.tif 1656 | river_331.tif 1657 | river_331.tif 1658 | river_331.tif 1659 | river_331.tif 1660 | river_331.tif 1661 | river_332.tif 1662 | river_332.tif 1663 | river_332.tif 1664 | river_332.tif 1665 | river_332.tif 1666 | river_333.tif 1667 | river_333.tif 1668 | river_333.tif 1669 | river_333.tif 1670 | river_333.tif 1671 | river_334.tif 1672 | river_334.tif 1673 | river_334.tif 1674 | river_334.tif 1675 | river_334.tif 1676 | river_335.tif 1677 | river_335.tif 1678 | river_335.tif 1679 | river_335.tif 1680 | river_335.tif 1681 | river_336.tif 1682 | river_336.tif 1683 | river_336.tif 1684 | river_336.tif 1685 | river_336.tif 1686 | school_337.tif 1687 | school_337.tif 1688 | school_337.tif 1689 | school_337.tif 1690 | school_337.tif 1691 | school_338.tif 1692 | school_338.tif 1693 | school_338.tif 1694 | school_338.tif 1695 | school_338.tif 1696 | school_339.tif 1697 | school_339.tif 1698 | school_339.tif 1699 | school_339.tif 1700 | school_339.tif 1701 | school_340.tif 1702 | school_340.tif 1703 | school_340.tif 1704 | school_340.tif 1705 | school_340.tif 1706 | school_341.tif 1707 | school_341.tif 1708 | school_341.tif 1709 | school_341.tif 1710 | school_341.tif 1711 | school_342.tif 1712 | school_342.tif 1713 | school_342.tif 1714 | school_342.tif 1715 | school_342.tif 1716 | school_343.tif 1717 | school_343.tif 1718 | school_343.tif 1719 | school_343.tif 1720 | school_343.tif 1721 | school_344.tif 1722 | school_344.tif 1723 | school_344.tif 1724 | school_344.tif 1725 | school_344.tif 1726 | school_345.tif 1727 | school_345.tif 1728 | school_345.tif 1729 | school_345.tif 1730 | school_345.tif 1731 | school_346.tif 1732 | school_346.tif 1733 | school_346.tif 1734 | school_346.tif 1735 | school_346.tif 1736 | school_347.tif 1737 | school_347.tif 1738 | school_347.tif 1739 | school_347.tif 1740 | school_347.tif 1741 | school_348.tif 1742 | school_348.tif 1743 | school_348.tif 1744 | school_348.tif 1745 | school_348.tif 1746 | school_349.tif 1747 | school_349.tif 1748 | school_349.tif 1749 | school_349.tif 1750 | school_349.tif 1751 | school_350.tif 1752 | school_350.tif 1753 | school_350.tif 1754 | school_350.tif 1755 | school_350.tif 1756 | school_351.tif 1757 | school_351.tif 1758 | school_351.tif 1759 | school_351.tif 1760 | school_351.tif 1761 | school_352.tif 1762 | school_352.tif 1763 | school_352.tif 1764 | school_352.tif 1765 | school_352.tif 1766 | sparseresidential_353.tif 1767 | sparseresidential_353.tif 1768 | sparseresidential_353.tif 1769 | sparseresidential_353.tif 1770 | sparseresidential_353.tif 1771 | sparseresidential_354.tif 1772 | sparseresidential_354.tif 1773 | sparseresidential_354.tif 1774 | sparseresidential_354.tif 1775 | sparseresidential_354.tif 1776 | sparseresidential_355.tif 1777 | sparseresidential_355.tif 1778 | sparseresidential_355.tif 1779 | sparseresidential_355.tif 1780 | sparseresidential_355.tif 1781 | sparseresidential_356.tif 1782 | sparseresidential_356.tif 1783 | sparseresidential_356.tif 1784 | sparseresidential_356.tif 1785 | sparseresidential_356.tif 1786 | sparseresidential_357.tif 1787 | sparseresidential_357.tif 1788 | sparseresidential_357.tif 1789 | sparseresidential_357.tif 1790 | sparseresidential_357.tif 1791 | sparseresidential_358.tif 1792 | sparseresidential_358.tif 1793 | sparseresidential_358.tif 1794 | sparseresidential_358.tif 1795 | sparseresidential_358.tif 1796 | sparseresidential_359.tif 1797 | sparseresidential_359.tif 1798 | sparseresidential_359.tif 1799 | sparseresidential_359.tif 1800 | sparseresidential_359.tif 1801 | sparseresidential_360.tif 1802 | sparseresidential_360.tif 1803 | sparseresidential_360.tif 1804 | sparseresidential_360.tif 1805 | sparseresidential_360.tif 1806 | sparseresidential_361.tif 1807 | sparseresidential_361.tif 1808 | sparseresidential_361.tif 1809 | sparseresidential_361.tif 1810 | sparseresidential_361.tif 1811 | sparseresidential_362.tif 1812 | sparseresidential_362.tif 1813 | sparseresidential_362.tif 1814 | sparseresidential_362.tif 1815 | sparseresidential_362.tif 1816 | sparseresidential_363.tif 1817 | sparseresidential_363.tif 1818 | sparseresidential_363.tif 1819 | sparseresidential_363.tif 1820 | sparseresidential_363.tif 1821 | sparseresidential_364.tif 1822 | sparseresidential_364.tif 1823 | sparseresidential_364.tif 1824 | sparseresidential_364.tif 1825 | sparseresidential_364.tif 1826 | sparseresidential_365.tif 1827 | sparseresidential_365.tif 1828 | sparseresidential_365.tif 1829 | sparseresidential_365.tif 1830 | sparseresidential_365.tif 1831 | sparseresidential_366.tif 1832 | sparseresidential_366.tif 1833 | sparseresidential_366.tif 1834 | sparseresidential_366.tif 1835 | sparseresidential_366.tif 1836 | sparseresidential_367.tif 1837 | sparseresidential_367.tif 1838 | sparseresidential_367.tif 1839 | sparseresidential_367.tif 1840 | sparseresidential_367.tif 1841 | square_368.tif 1842 | square_368.tif 1843 | square_368.tif 1844 | square_368.tif 1845 | square_368.tif 1846 | square_369.tif 1847 | square_369.tif 1848 | square_369.tif 1849 | square_369.tif 1850 | square_369.tif 1851 | square_370.tif 1852 | square_370.tif 1853 | square_370.tif 1854 | square_370.tif 1855 | square_370.tif 1856 | square_371.tif 1857 | square_371.tif 1858 | square_371.tif 1859 | square_371.tif 1860 | square_371.tif 1861 | square_372.tif 1862 | square_372.tif 1863 | square_372.tif 1864 | square_372.tif 1865 | square_372.tif 1866 | square_373.tif 1867 | square_373.tif 1868 | square_373.tif 1869 | square_373.tif 1870 | square_373.tif 1871 | square_374.tif 1872 | square_374.tif 1873 | square_374.tif 1874 | square_374.tif 1875 | square_374.tif 1876 | square_375.tif 1877 | square_375.tif 1878 | square_375.tif 1879 | square_375.tif 1880 | square_375.tif 1881 | square_376.tif 1882 | square_376.tif 1883 | square_376.tif 1884 | square_376.tif 1885 | square_376.tif 1886 | square_377.tif 1887 | square_377.tif 1888 | square_377.tif 1889 | square_377.tif 1890 | square_377.tif 1891 | square_378.tif 1892 | square_378.tif 1893 | square_378.tif 1894 | square_378.tif 1895 | square_378.tif 1896 | square_379.tif 1897 | square_379.tif 1898 | square_379.tif 1899 | square_379.tif 1900 | square_379.tif 1901 | square_380.tif 1902 | square_380.tif 1903 | square_380.tif 1904 | square_380.tif 1905 | square_380.tif 1906 | square_381.tif 1907 | square_381.tif 1908 | square_381.tif 1909 | square_381.tif 1910 | square_381.tif 1911 | square_382.tif 1912 | square_382.tif 1913 | square_382.tif 1914 | square_382.tif 1915 | square_382.tif 1916 | square_383.tif 1917 | square_383.tif 1918 | square_383.tif 1919 | square_383.tif 1920 | square_383.tif 1921 | square_384.tif 1922 | square_384.tif 1923 | square_384.tif 1924 | square_384.tif 1925 | square_384.tif 1926 | square_385.tif 1927 | square_385.tif 1928 | square_385.tif 1929 | square_385.tif 1930 | square_385.tif 1931 | square_386.tif 1932 | square_386.tif 1933 | square_386.tif 1934 | square_386.tif 1935 | square_386.tif 1936 | stadium_387.tif 1937 | stadium_387.tif 1938 | stadium_387.tif 1939 | stadium_387.tif 1940 | stadium_387.tif 1941 | stadium_388.tif 1942 | stadium_388.tif 1943 | stadium_388.tif 1944 | stadium_388.tif 1945 | stadium_388.tif 1946 | stadium_389.tif 1947 | stadium_389.tif 1948 | stadium_389.tif 1949 | stadium_389.tif 1950 | stadium_389.tif 1951 | stadium_390.tif 1952 | stadium_390.tif 1953 | stadium_390.tif 1954 | stadium_390.tif 1955 | stadium_390.tif 1956 | stadium_391.tif 1957 | stadium_391.tif 1958 | stadium_391.tif 1959 | stadium_391.tif 1960 | stadium_391.tif 1961 | stadium_392.tif 1962 | stadium_392.tif 1963 | stadium_392.tif 1964 | stadium_392.tif 1965 | stadium_392.tif 1966 | stadium_393.tif 1967 | stadium_393.tif 1968 | stadium_393.tif 1969 | stadium_393.tif 1970 | stadium_393.tif 1971 | stadium_394.tif 1972 | stadium_394.tif 1973 | stadium_394.tif 1974 | stadium_394.tif 1975 | stadium_394.tif 1976 | stadium_395.tif 1977 | stadium_395.tif 1978 | stadium_395.tif 1979 | stadium_395.tif 1980 | stadium_395.tif 1981 | stadium_396.tif 1982 | stadium_396.tif 1983 | stadium_396.tif 1984 | stadium_396.tif 1985 | stadium_396.tif 1986 | stadium_397.tif 1987 | stadium_397.tif 1988 | stadium_397.tif 1989 | stadium_397.tif 1990 | stadium_397.tif 1991 | stadium_398.tif 1992 | stadium_398.tif 1993 | stadium_398.tif 1994 | stadium_398.tif 1995 | stadium_398.tif 1996 | stadium_399.tif 1997 | stadium_399.tif 1998 | stadium_399.tif 1999 | stadium_399.tif 2000 | stadium_399.tif 2001 | stadium_400.tif 2002 | stadium_400.tif 2003 | stadium_400.tif 2004 | stadium_400.tif 2005 | stadium_400.tif 2006 | stadium_401.tif 2007 | stadium_401.tif 2008 | stadium_401.tif 2009 | stadium_401.tif 2010 | stadium_401.tif 2011 | stadium_402.tif 2012 | stadium_402.tif 2013 | stadium_402.tif 2014 | stadium_402.tif 2015 | stadium_402.tif 2016 | stadium_403.tif 2017 | stadium_403.tif 2018 | stadium_403.tif 2019 | stadium_403.tif 2020 | stadium_403.tif 2021 | stadium_404.tif 2022 | stadium_404.tif 2023 | stadium_404.tif 2024 | stadium_404.tif 2025 | stadium_404.tif 2026 | stadium_405.tif 2027 | stadium_405.tif 2028 | stadium_405.tif 2029 | stadium_405.tif 2030 | stadium_405.tif 2031 | storagetanks_406.tif 2032 | storagetanks_406.tif 2033 | storagetanks_406.tif 2034 | storagetanks_406.tif 2035 | storagetanks_406.tif 2036 | storagetanks_407.tif 2037 | storagetanks_407.tif 2038 | storagetanks_407.tif 2039 | storagetanks_407.tif 2040 | storagetanks_407.tif 2041 | storagetanks_408.tif 2042 | storagetanks_408.tif 2043 | storagetanks_408.tif 2044 | storagetanks_408.tif 2045 | storagetanks_408.tif 2046 | storagetanks_409.tif 2047 | storagetanks_409.tif 2048 | storagetanks_409.tif 2049 | storagetanks_409.tif 2050 | storagetanks_409.tif 2051 | storagetanks_410.tif 2052 | storagetanks_410.tif 2053 | storagetanks_410.tif 2054 | storagetanks_410.tif 2055 | storagetanks_410.tif 2056 | storagetanks_411.tif 2057 | storagetanks_411.tif 2058 | storagetanks_411.tif 2059 | storagetanks_411.tif 2060 | storagetanks_411.tif 2061 | storagetanks_412.tif 2062 | storagetanks_412.tif 2063 | storagetanks_412.tif 2064 | storagetanks_412.tif 2065 | storagetanks_412.tif 2066 | storagetanks_413.tif 2067 | storagetanks_413.tif 2068 | storagetanks_413.tif 2069 | storagetanks_413.tif 2070 | storagetanks_413.tif 2071 | storagetanks_414.tif 2072 | storagetanks_414.tif 2073 | storagetanks_414.tif 2074 | storagetanks_414.tif 2075 | storagetanks_414.tif 2076 | storagetanks_415.tif 2077 | storagetanks_415.tif 2078 | storagetanks_415.tif 2079 | storagetanks_415.tif 2080 | storagetanks_415.tif 2081 | storagetanks_416.tif 2082 | storagetanks_416.tif 2083 | storagetanks_416.tif 2084 | storagetanks_416.tif 2085 | storagetanks_416.tif 2086 | storagetanks_417.tif 2087 | storagetanks_417.tif 2088 | storagetanks_417.tif 2089 | storagetanks_417.tif 2090 | storagetanks_417.tif 2091 | storagetanks_418.tif 2092 | storagetanks_418.tif 2093 | storagetanks_418.tif 2094 | storagetanks_418.tif 2095 | storagetanks_418.tif 2096 | storagetanks_419.tif 2097 | storagetanks_419.tif 2098 | storagetanks_419.tif 2099 | storagetanks_419.tif 2100 | storagetanks_419.tif 2101 | storagetanks_420.tif 2102 | storagetanks_420.tif 2103 | storagetanks_420.tif 2104 | storagetanks_420.tif 2105 | storagetanks_420.tif 2106 | storagetanks_421.tif 2107 | storagetanks_421.tif 2108 | storagetanks_421.tif 2109 | storagetanks_421.tif 2110 | storagetanks_421.tif 2111 | storagetanks_422.tif 2112 | storagetanks_422.tif 2113 | storagetanks_422.tif 2114 | storagetanks_422.tif 2115 | storagetanks_422.tif 2116 | storagetanks_423.tif 2117 | storagetanks_423.tif 2118 | storagetanks_423.tif 2119 | storagetanks_423.tif 2120 | storagetanks_423.tif 2121 | storagetanks_424.tif 2122 | storagetanks_424.tif 2123 | storagetanks_424.tif 2124 | storagetanks_424.tif 2125 | storagetanks_424.tif 2126 | storagetanks_425.tif 2127 | storagetanks_425.tif 2128 | storagetanks_425.tif 2129 | storagetanks_425.tif 2130 | storagetanks_425.tif 2131 | storagetanks_426.tif 2132 | storagetanks_426.tif 2133 | storagetanks_426.tif 2134 | storagetanks_426.tif 2135 | storagetanks_426.tif 2136 | storagetanks_427.tif 2137 | storagetanks_427.tif 2138 | storagetanks_427.tif 2139 | storagetanks_427.tif 2140 | storagetanks_427.tif 2141 | storagetanks_428.tif 2142 | storagetanks_428.tif 2143 | storagetanks_428.tif 2144 | storagetanks_428.tif 2145 | storagetanks_428.tif 2146 | viaduct_429.tif 2147 | viaduct_429.tif 2148 | viaduct_429.tif 2149 | viaduct_429.tif 2150 | viaduct_429.tif 2151 | viaduct_430.tif 2152 | viaduct_430.tif 2153 | viaduct_430.tif 2154 | viaduct_430.tif 2155 | viaduct_430.tif 2156 | viaduct_431.tif 2157 | viaduct_431.tif 2158 | viaduct_431.tif 2159 | viaduct_431.tif 2160 | viaduct_431.tif 2161 | viaduct_432.tif 2162 | viaduct_432.tif 2163 | viaduct_432.tif 2164 | viaduct_432.tif 2165 | viaduct_432.tif 2166 | viaduct_433.tif 2167 | viaduct_433.tif 2168 | viaduct_433.tif 2169 | viaduct_433.tif 2170 | viaduct_433.tif 2171 | viaduct_434.tif 2172 | viaduct_434.tif 2173 | viaduct_434.tif 2174 | viaduct_434.tif 2175 | viaduct_434.tif 2176 | viaduct_435.tif 2177 | viaduct_435.tif 2178 | viaduct_435.tif 2179 | viaduct_435.tif 2180 | viaduct_435.tif 2181 | viaduct_436.tif 2182 | viaduct_436.tif 2183 | viaduct_436.tif 2184 | viaduct_436.tif 2185 | viaduct_436.tif 2186 | viaduct_437.tif 2187 | viaduct_437.tif 2188 | viaduct_437.tif 2189 | viaduct_437.tif 2190 | viaduct_437.tif 2191 | viaduct_438.tif 2192 | viaduct_438.tif 2193 | viaduct_438.tif 2194 | viaduct_438.tif 2195 | viaduct_438.tif 2196 | viaduct_439.tif 2197 | viaduct_439.tif 2198 | viaduct_439.tif 2199 | viaduct_439.tif 2200 | viaduct_439.tif 2201 | viaduct_440.tif 2202 | viaduct_440.tif 2203 | viaduct_440.tif 2204 | viaduct_440.tif 2205 | viaduct_440.tif 2206 | viaduct_441.tif 2207 | viaduct_441.tif 2208 | viaduct_441.tif 2209 | viaduct_441.tif 2210 | viaduct_441.tif 2211 | viaduct_442.tif 2212 | viaduct_442.tif 2213 | viaduct_442.tif 2214 | viaduct_442.tif 2215 | viaduct_442.tif 2216 | viaduct_443.tif 2217 | viaduct_443.tif 2218 | viaduct_443.tif 2219 | viaduct_443.tif 2220 | viaduct_443.tif 2221 | viaduct_444.tif 2222 | viaduct_444.tif 2223 | viaduct_444.tif 2224 | viaduct_444.tif 2225 | viaduct_444.tif 2226 | viaduct_445.tif 2227 | viaduct_445.tif 2228 | viaduct_445.tif 2229 | viaduct_445.tif 2230 | viaduct_445.tif 2231 | viaduct_446.tif 2232 | viaduct_446.tif 2233 | viaduct_446.tif 2234 | viaduct_446.tif 2235 | viaduct_446.tif 2236 | denseresidential_447.tif 2237 | denseresidential_447.tif 2238 | denseresidential_447.tif 2239 | denseresidential_447.tif 2240 | denseresidential_447.tif 2241 | airport_448.tif 2242 | airport_448.tif 2243 | airport_448.tif 2244 | airport_448.tif 2245 | airport_448.tif 2246 | intersection_449.tif 2247 | intersection_449.tif 2248 | intersection_449.tif 2249 | intersection_449.tif 2250 | intersection_449.tif 2251 | playground_450.tif 2252 | playground_450.tif 2253 | playground_450.tif 2254 | playground_450.tif 2255 | playground_450.tif 2256 | storagetanks_451.tif 2257 | storagetanks_451.tif 2258 | storagetanks_451.tif 2259 | storagetanks_451.tif 2260 | storagetanks_451.tif 2261 | -------------------------------------------------------------------------------- /predict/evaluate_selo.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/05 6 | 7 | import json 8 | import logging 9 | import os 10 | import sys 11 | 12 | sys.path.append("..") 13 | from evaluations.SLM import SLM 14 | 15 | if __name__ == "__main__": 16 | import argparse 17 | 18 | # settings 19 | parser = argparse.ArgumentParser(description="SLM") 20 | parser.add_argument("--yaml_path", type=str, default="option/RSITMD/RSITMD_AMFMN.yaml", help="config yaml path") 21 | parser.add_argument("--cache_path", type=str, default="cache/RSITMD_AMFMN", help="cache path") 22 | parser.add_argument("--src_data_path", type=str, default="../test_data/imgs", help="testset images path") 23 | parser.add_argument("--src_anno_path", type=str, default="../test_data/annotations/anno.json", help="testset annotations path") 24 | opt = parser.parse_args() 25 | 26 | # logging 27 | logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s') 28 | logger = logging.getLogger(__name__) 29 | 30 | # start eval 31 | slm_metric = SLM() 32 | 33 | # load from annotations 34 | with open(opt.src_anno_path,'r',encoding='utf8')as fp: 35 | json_data = json.load(fp) 36 | 37 | for idx, item in enumerate(json_data): 38 | # load sample 39 | img = item['jpg_name'] 40 | text = item['caption'] 41 | probmap_path = os.path.join(opt.cache_path, "probmap_{}.jpg".format(idx)) 42 | points = item['points'] 43 | 44 | # logging 45 | logger.info("Processing {}/{}: {}".format(idx, len(json_data), img)) 46 | logger.info("Corresponding text: {}".format(text)) 47 | 48 | # evaluate # 49 | metrics = slm_metric.evaluate(probmap_path, region_list=points) 50 | slm_metric.append_metric(metrics) 51 | 52 | slm_metric.get_the_mean_metric() -------------------------------------------------------------------------------- /predict/generate_selo.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/03 6 | 7 | import json 8 | import os 9 | import sys 10 | import time 11 | 12 | import cv2 13 | import numpy as np 14 | import utils 15 | from model_encoder import Encoder 16 | from model_init import model_init 17 | 18 | sys.path.append("..") 19 | from evaluations.SLM import SLM 20 | 21 | def split_image(img_path, steps, cache_path): 22 | t1 = time.time() 23 | 24 | subimage_files_dir = os.path.join(cache_path, os.path.basename(img_path).split(".")[0]) 25 | 26 | # 裁切图像文件夹 27 | subimages_dir = subimage_files_dir +'_subimages' 28 | if os.path.exists(subimages_dir): 29 | utils.delete_dire(subimages_dir) 30 | else: 31 | os.makedirs(subimages_dir) 32 | 33 | # Read Image 34 | source_img = cv2.imread(img_path) 35 | img_weight = np.shape(source_img)[0] 36 | img_height = np.shape(source_img)[1] 37 | logger.info("img size:{}x{}".format(img_weight, img_height)) 38 | 39 | for step in steps: 40 | logger.info("Start split images with step {}".format(step)) 41 | for gap in [step, 0.5 * step]: 42 | gap = int(gap) 43 | 44 | # Cut img 45 | for h in range(0 + (step - gap), img_height, step): 46 | h_start, h_end = h, h + step 47 | # bound? 48 | if h_end >= img_height: 49 | h_start, h_end = img_height - step, img_height 50 | 51 | for w in range(0 + (step - gap), img_weight, step): 52 | w_start, w_end = w, w + step 53 | # bound? 54 | if w_end >= img_weight: 55 | w_start, w_end = img_weight - step, img_weight 56 | 57 | cut_img_name = str(w_start) + "_" + str(w_end) + "_" + str(h_start) + "_" + str(h_end) + ".jpg" 58 | cut_img = source_img[w_start:w_end, h_start:h_end] 59 | cut_img = cv2.resize(cut_img, (256, 256), interpolation=cv2.INTER_CUBIC) 60 | 61 | cv2.imwrite(os.path.join(subimages_dir, cut_img_name), cut_img) 62 | 63 | 64 | logger.info("Image {} has been split successfully.".format(img_path)) 65 | 66 | return time.time() - t1 67 | 68 | def generate_heatmap(img_path, text, output_file_h, output_file_a, output_file_p, cache_path): 69 | 70 | subimages_dir = os.path.join(cache_path, os.path.basename(img_path).split(".")[0]) +'_subimages' 71 | 72 | logger.info("Start calculate similarities ...") 73 | cal_start = time.time() 74 | 75 | # init encoder 76 | encoder = Encoder(model) 77 | 78 | # text vector 79 | text_vector = encoder.text_encoder(model, vocab, text) 80 | 81 | # read subimages 82 | subimages = os.listdir(subimages_dir) 83 | sim_results = [] 84 | for subimage in subimages: 85 | image_vector = encoder.image_encoder(model, os.path.join(subimages_dir, subimage)) 86 | sim_results.append(encoder.cosine_sim(text_vector, image_vector)) 87 | cal_end = time.time() 88 | logger.info("Calculate similarities in {}s".format(cal_end-cal_start)) 89 | t2 = cal_end-cal_start 90 | 91 | logger.info("Start generate heatmap ...") 92 | generate_start = time.time() 93 | 94 | # read Image 95 | source_img = cv2.imread(img_path) 96 | img_row = np.shape(source_img)[0] 97 | img_col = np.shape(source_img)[1] 98 | 99 | # mkdir map 100 | heat_map = np.zeros([img_row, img_col], dtype=float) 101 | heat_num = np.zeros([img_row, img_col], dtype=float) 102 | for idx,file in enumerate(subimages): 103 | r_start, r_end, c_start, c_end = file.replace(".jpg","").split("_") 104 | 105 | heat_map[int(r_start):int(r_end), int(c_start):int(c_end)] += sim_results[idx] 106 | heat_num[int(r_start):int(r_end), int(c_start):int(c_end)] += 1 107 | 108 | 109 | for i in range(np.shape(heat_map)[0]): 110 | for j in range(np.shape(heat_map)[1]): 111 | heat_map[i,j] = heat_map[i,j] / heat_num[i,j] 112 | t3 = time.time() - generate_start 113 | 114 | logger.info("Generate finished, start optim ...") 115 | optim_start = time.time() 116 | # filter 117 | adaptive = np.asarray(heat_map) 118 | adaptive = adaptive-np.min(adaptive) 119 | probmap = adaptive/np.max(adaptive) 120 | # must convert to type unit8 121 | probmap = np.uint8(255 * probmap) 122 | probmap = cv2.medianBlur(probmap,251) 123 | heatmap = cv2.applyColorMap(probmap, cv2.COLORMAP_JET) 124 | img_add = cv2.addWeighted(source_img, 0.7, heatmap, 0.3, 0) 125 | generate_end = time.time() 126 | logger.info("Generate heatmap in {}s".format(generate_end-generate_start)) 127 | 128 | logger.info("Saving heatmap in {} ...".format(output_file_h)) 129 | logger.info("Saving heatmap in {} ...".format(output_file_a)) 130 | logger.info("Saving heatmap in {} ...".format(output_file_p)) 131 | cv2.imwrite( output_file_p ,probmap) 132 | cv2.imwrite( output_file_h ,heatmap) 133 | cv2.imwrite( output_file_a ,img_add) 134 | logger.info("Saved ok.") 135 | 136 | # clear temp 137 | utils.delete_dire(subimages_dir) 138 | os.rmdir(subimages_dir) 139 | 140 | t4 = generate_end - optim_start 141 | return t2, t3, t4 142 | 143 | if __name__ == "__main__": 144 | 145 | import argparse 146 | 147 | # settings 148 | parser = argparse.ArgumentParser(description="SLM") 149 | parser.add_argument("--yaml_path", type=str, default="option/RSITMD/RSITMD_AMFMN.yaml", help="config yaml path") 150 | parser.add_argument("--cache_path", type=str, default="cache/RSITMD_AMFMN", help="cache path") 151 | parser.add_argument("--src_data_path", type=str, default="../test_data/imgs", help="testset images path") 152 | parser.add_argument("--src_anno_path", type=str, default="../test_data/annotations/anno.json", help="testset annotations path") 153 | parser.add_argument("--step", type=str, default="256_512_768", help="step") 154 | opt = parser.parse_args() 155 | 156 | # mkdir 157 | if not os.path.exists(opt.cache_path): 158 | os.mkdir(opt.cache_path) 159 | 160 | # params 161 | steps = [int(step) for step in opt.step.split("_")] 162 | 163 | # logging 164 | logger = utils.get_logger(os.path.join(opt.cache_path, 'log.txt')) 165 | 166 | # init model 167 | model, vocab = model_init( 168 | prefix_path = "./", 169 | yaml_path = opt.yaml_path 170 | ) 171 | 172 | # start eval 173 | slm_metric = SLM() 174 | 175 | # load from annotations 176 | with open(opt.src_anno_path,'r',encoding='utf8')as fp: 177 | json_data = json.load(fp) 178 | 179 | t1_all, t2_all, t3_all, t4_all = 0, 0, 0, 0 180 | total_time = time.time() 181 | 182 | for idx, item in enumerate(json_data): 183 | # load sample 184 | img = item['jpg_name'] 185 | text = item['caption'] 186 | points = item['points'] 187 | 188 | # path 189 | img_path = os.path.join(opt.src_data_path, img) 190 | heatmap_path = os.path.join(opt.cache_path, "heatmap_{}.jpg".format(idx)) 191 | probmap_path = os.path.join(opt.cache_path, "probmap_{}.jpg".format(idx)) 192 | addmap_path = os.path.join(opt.cache_path, "addmap_{}.jpg".format(idx)) 193 | 194 | # logging 195 | logger.info("Processing {}/{}: {}".format(idx, len(json_data), img)) 196 | logger.info("Corresponding text: {}".format(text)) 197 | 198 | # processing 199 | t1 = split_image(img_path, steps, opt.cache_path) 200 | t2, t3, t4 = generate_heatmap(img_path, text, heatmap_path, addmap_path, probmap_path, opt.cache_path) 201 | 202 | t1_all += t1 203 | t2_all += t2 204 | t3_all += t3 205 | t4_all += t4 206 | 207 | # evaluate # 208 | metrics = slm_metric.evaluate(probmap_path, region_list=points) 209 | slm_metric.append_metric(metrics) 210 | 211 | slm_metric.get_the_mean_metric() 212 | 213 | all_time = time.time() - total_time 214 | logger.info("Time-Cut: {:.4f}s".format(t1_all)) 215 | logger.info("Time-Sim: {:.4f}".format(t2_all)) 216 | logger.info("Time-Gnt: {:.4f}".format(t3_all)) 217 | logger.info("Time-Flt: {:.4f}".format(t4_all)) 218 | logger.info("Time-Total: {:.4f}".format(all_time)) 219 | 220 | -------------------------------------------------------------------------------- /predict/generate_selo_v2.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/03 6 | 7 | import json 8 | import os 9 | import sys 10 | import time 11 | 12 | import cv2 13 | import numpy as np 14 | 15 | import utils 16 | from model_encoder import Encoder 17 | from model_init import model_init 18 | 19 | sys.path.append("..") 20 | from evaluations.SLM import SLM 21 | 22 | # 将图片按照512*512剪裁，并返回剪裁时间和剪切次数 23 | # img_path：图片位置，比如0.jpg 24 | # subimages_dir：剪裁的图片放的位置 25 | # cut_count：剪裁的次数 26 | def split_image(img_path, subimages_dir): 27 | t1 = time.time() 28 | # Read Image 29 | source_img = cv2.imread(img_path) 30 | img_height = np.shape(source_img)[0] 31 | img_weight = np.shape(source_img)[1] 32 | step = 512 33 | # 剪裁图像 34 | # Cut img 35 | for h in range(0, img_height, step): 36 | h_start, h_end = h, h + step 37 | if h_end >= img_height: 38 | h_start, h_end = img_height - step, img_height 39 | for w in range(0, img_weight, step): 40 | w_start, w_end = w, w + step 41 | # bound? 42 | if w_end >= img_weight: 43 | w_start, w_end = img_weight - step, img_weight 44 | cut_img_name = str(h_start) + "_" + str(h_end) + "_" + str(w_start) + "_" + str(w_end) + ".jpg" 45 | cut_img = source_img[h_start:h_end, w_start:w_end] 46 | cut_img = cv2.resize(cut_img, (256, 256), interpolation=cv2.INTER_CUBIC) 47 | cv2.imwrite(os.path.join(subimages_dir, cut_img_name), cut_img) 48 | split_time = time.time() - t1 49 | 50 | return split_time 51 | 52 | # 获取裁剪图像的特征向量 53 | # subimages_dir：剪裁好的图片所在位置，比如：0_subimages 54 | # img_vectors：每一张剪裁好图片的特征向量的列表集合 55 | # n_clusters：k-means的聚类个数 56 | # text：文本 57 | def get_img_vectors(subimages_dir, text, boximages_dir, encoder, img_path, sim_results1): 58 | 59 | t2 = time.time() 60 | 61 | # read subimages 62 | subimages = os.listdir(subimages_dir) 63 | 64 | # 获取文本的特征向量 65 | text_vector = encoder.text_encoder(model, vocab, text) 66 | 67 | random_times = 4 68 | 69 | for i in range(random_times): 70 | tmp = text.split() 71 | np.random.shuffle(tmp) 72 | text_bak = " ".join([i for i in tmp]) 73 | text_vector += encoder.text_encoder(model, vocab, text_bak) 74 | text_vector /= (random_times + 1) 75 | 76 | # 获取每一个剪裁好图片的特征,并计算与文本的相似度 77 | for subimage in subimages: 78 | # 一维的特征向量 79 | image_vector = encoder.image_encoder(model, os.path.join(subimages_dir, subimage)) 80 | sim = encoder.cosine_sim(image_vector,text_vector) 81 | sim_results1.append(sim) 82 | 83 | 84 | # 将相似度排序，返回索引列表 85 | sim_indexs = np.argsort(sim_results1) 86 | source_img = cv2.imread(img_path) 87 | img_height = np.shape(source_img)[0] 88 | img_weight = np.shape(source_img)[1] 89 | 90 | 91 | select1 = int(len(sim_indexs)*0.15)+1 92 | select_index1 = sim_indexs[-select1:] 93 | 94 | for i in select_index1: 95 | img_name = subimages[i] 96 | h_start, h_end, w_start, w_end = img_name.replace(".jpg", "").split("_") 97 | cut_img_size = 512 98 | p1 = [int(h_start), int(w_start)] 99 | p2 = [int(h_start), int(w_end)] 100 | p3 = [int(h_end), int(w_start)] 101 | p4 = [int(h_end), int(w_end)] 102 | p5 = [int(h_start) + cut_img_size * 1 / 4, int(w_start) + cut_img_size * 1 / 4] 103 | p6 = [int(h_start) + cut_img_size * 1 / 4, int(w_start) + cut_img_size * 3 / 4] 104 | p7 = [int(h_start) + cut_img_size * 3 / 4, int(w_start) + cut_img_size * 1 / 4] 105 | p8 = [int(h_start) + cut_img_size * 3 / 4, int(w_start) + cut_img_size * 3 / 4] 106 | p9 = [int(h_start) + cut_img_size * 1 / 2, int(w_start) + cut_img_size * 1 / 2] 107 | points = [p1, p2, p3, p4, p5, p6, p7, p8, p9] 108 | for point in points: 109 | h = point[0] 110 | w = point[1] 111 | sizes = [64,128] 112 | for size in sizes: 113 | h_start1 = int(h - size if h - size > 0 else 0) 114 | h_end1 = int(h_start1 + size * 2 if h_start1 + size * 2 < img_height else img_height) 115 | w_start1 = int(w - size if w - size > 0 else 0) 116 | w_end1 = int(w_start1 + size * 2 if w_start1 + size * 2 < img_weight else img_weight) 117 | box_img_name = str(h_start1) + "_" + str(h_end1) + "_" + str(w_start1) + "_" + str(w_end1) + ".jpg" 118 | box_img = source_img[h_start1:h_end1, w_start1:w_end1] 119 | box_img = cv2.resize(box_img, (256, 256), interpolation=cv2.INTER_CUBIC) 120 | cv2.imwrite(os.path.join(boximages_dir, box_img_name), box_img) 121 | stage2_time = time.time()-t2 122 | 123 | 124 | return text_vector,select_index1,stage2_time 125 | 126 | def generate_heatmap(img_path, output_file_h, output_file_a, output_file_p, text_vector, boximages_dir, subimages_dir,sim_results1,select_index1, 127 | encoder): 128 | 129 | t3 = time.time() 130 | 131 | # read subimages 132 | boximages = os.listdir(boximages_dir) 133 | subimages = os.listdir(subimages_dir) 134 | 135 | sim_results = [] 136 | for boximage in boximages: 137 | image_vector = encoder.image_encoder(model, os.path.join(boximages_dir, boximage)) 138 | sim_results.append(encoder.cosine_sim(text_vector, image_vector)) 139 | 140 | 141 | # read Image 142 | source_img = cv2.imread(img_path) 143 | img_row = np.shape(source_img)[0] 144 | img_col = np.shape(source_img)[1] 145 | 146 | # mkdir map 147 | heat_map = np.zeros([img_row, img_col], dtype=float) 148 | heat_num = np.zeros([img_row, img_col], dtype=float) 149 | 150 | for idx in select_index1: 151 | h_start, h_end, w_start, w_end = subimages[idx].replace(".jpg", "").split("_") 152 | heat_map[int(h_start):int(h_end), int(w_start):int(w_end)] += sim_results1[idx] 153 | heat_num[int(h_start):int(h_end), int(w_start):int(w_end)] += 1 154 | 155 | 156 | for idx,file in enumerate(boximages): 157 | h_start, h_end, w_start, w_end = file.replace(".jpg","").split("_") 158 | heat_map[int(h_start):int(h_end), int(w_start):int(w_end)] += sim_results[idx] 159 | heat_num[int(h_start):int(h_end), int(w_start):int(w_end)] += 1 160 | 161 | for i in range(np.shape(heat_map)[0]): 162 | for j in range(np.shape(heat_map)[1]): 163 | if(heat_num[i,j]!=0.0): 164 | heat_map[i,j] = heat_map[i,j] / heat_num[i,j] 165 | 166 | # filter 167 | adaptive = np.asarray(heat_map) 168 | adaptive = adaptive-np.min(adaptive) 169 | probmap = adaptive/np.max(adaptive) 170 | # must convert to type unit8 171 | probmap = np.uint8(255 * probmap) 172 | # 中值滤波 173 | # probmap = cv2.medianBlur(probmap,251) 174 | # 高斯滤波 175 | probmap = cv2.GaussianBlur(probmap, (255, 255), 0, 0) 176 | heatmap = cv2.applyColorMap(probmap, cv2.COLORMAP_JET) 177 | img_add = cv2.addWeighted(source_img, 0.7, heatmap, 0.3, 0) 178 | t3 = time.time()-t3 179 | 180 | cv2.imwrite( output_file_p ,probmap) 181 | cv2.imwrite( output_file_h ,heatmap) 182 | cv2.imwrite( output_file_a ,img_add) 183 | 184 | # clear temp 185 | utils.delete_dire(boximages_dir) 186 | os.rmdir(boximages_dir) 187 | 188 | utils.delete_dire(subimages_dir) 189 | os.rmdir(subimages_dir) 190 | 191 | return t3 192 | 193 | 194 | if __name__ == "__main__": 195 | 196 | import argparse 197 | 198 | # settings 199 | parser = argparse.ArgumentParser(description="SLM") 200 | parser.add_argument("--yaml_path", type=str, default="option/RSITMD/RSITMD_AMFMN.yaml", help="config yaml path") 201 | parser.add_argument("--cache_path", type=str, default="cache/RSITMD_AMFMN", help="cache path") 202 | parser.add_argument("--src_data_path", type=str, default="../test_data/imgs", help="testset images path") 203 | parser.add_argument("--src_anno_path", type=str, default="../test_data/annotations/anno.json", 204 | help="testset annotations path") 205 | opt = parser.parse_args() 206 | 207 | # mkdir 208 | if not os.path.exists(opt.cache_path): 209 | os.mkdir(opt.cache_path) 210 | 211 | # logging 212 | logger = utils.get_logger(os.path.join(opt.cache_path, 'log.txt')) 213 | 214 | # init model 215 | model, vocab = model_init( 216 | prefix_path="./", 217 | yaml_path=opt.yaml_path 218 | ) 219 | 220 | # init encoder 221 | encoder = Encoder(model) 222 | 223 | # start eval 224 | slm_metric = SLM() 225 | 226 | # load from annotations 227 | with open(opt.src_anno_path, 'r', encoding='utf8')as fp: 228 | json_data = json.load(fp) 229 | 230 | total_time = time.time() 231 | second_time_list = [] 232 | 233 | for idx, item in enumerate(json_data): 234 | # load sample 235 | img = item['jpg_name'] 236 | text = item['caption'] 237 | points = item['points'] 238 | count = 0 239 | 240 | # path 241 | img_path = os.path.join(opt.src_data_path, img) 242 | subimage_files_dir = os.path.join(opt.cache_path, os.path.basename(img_path).split(".")[0]) 243 | heatmap_path = os.path.join(opt.cache_path, "heatmap_{}.jpg".format(idx)) 244 | probmap_path = os.path.join(opt.cache_path, "probmap_{}.jpg".format(idx)) 245 | addmap_path = os.path.join(opt.cache_path, "addmap_{}.jpg".format(idx)) 246 | 247 | # 裁切图像文件夹 248 | subimages_dir = subimage_files_dir + '_subimages' 249 | # anchor_box的文件夹 250 | boximages_dir = subimage_files_dir + '_boximages' 251 | if os.path.exists(subimages_dir): 252 | utils.delete_dire(subimages_dir) 253 | else: 254 | os.makedirs(subimages_dir) 255 | 256 | if (os.path.exists(boximages_dir)): 257 | utils.delete_dire(boximages_dir) 258 | else: 259 | os.makedirs(boximages_dir) 260 | 261 | times = time.time() 262 | 263 | # processing 264 | split_image(img_path, subimages_dir) 265 | 266 | sim_results1 = [] 267 | select_index1 = [] 268 | stage2_time = 0 269 | text_vector, select_index1, stage2_time = get_img_vectors(subimages_dir, text, boximages_dir, encoder, img_path, 270 | sim_results1) 271 | 272 | generate_heatmap(img_path, heatmap_path, addmap_path, probmap_path, text_vector, 273 | boximages_dir, subimages_dir, sim_results1, select_index1, encoder) 274 | 275 | times = time.time() - times 276 | second_time_list.append(times) 277 | logger.info("Processing {}/{}: {}，的总时间为：{}".format(idx, len(json_data), img, times)) 278 | logger.info("Corresponding text: {}".format(text)) 279 | 280 | # evaluate # 281 | metrics = slm_metric.evaluate(probmap_path, region_list=points) 282 | slm_metric.append_metric(metrics) 283 | 284 | slm_metric.get_the_mean_metric() 285 | 286 | all_time = time.time() - total_time 287 | logger.info("Time-Total: {:.4f}".format(all_time)) 288 | logger.info("second_time_list={}".format(second_time_list)) 289 | -------------------------------------------------------------------------------- /predict/layers/AMFMN.py: -------------------------------------------------------------------------------- 1 | # ----------------------------------------------------------- 2 | # "Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information" 3 | # Yuan, Zhiqiang and Zhang, Wenkai and Changyuan Tian and Xuee, Rong and Zhengyuan Zhang and Wang, Hongqi and Fu, Kun and Sun, Xian 4 | # Writen by YuanZhiqiang, 2021. Our code is depended on AMFMN 5 | # ------------------------------------------------------------ 6 | import copy 7 | 8 | import torch.nn.init 9 | 10 | from .AMFMN_utils import * 11 | 12 | 13 | def l2norm(X, dim, eps=1e-8): 14 | """L2-normalize columns of X 15 | """ 16 | norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps 17 | X = torch.div(X, norm) 18 | return X 19 | 20 | def cosine_sim(im, s): 21 | """Cosine similarity between all the image and sentence pairs 22 | """ 23 | im = l2norm(im, dim=-1) 24 | s = l2norm(s, dim=-1) 25 | w12 = im.mm(s.t()) 26 | return w12 27 | 28 | class BaseModel(nn.Module): 29 | def __init__(self, opt={}, vocab_words=[]): 30 | super(BaseModel, self).__init__() 31 | 32 | # img feature 33 | self.extract_feature = ExtractFeature(opt = opt) 34 | self.drop_g_v = nn.Dropout(0.3) 35 | 36 | # vsa feature 37 | self.mvsa =VSA_Module(opt = opt) 38 | 39 | # text feature 40 | self.text_feature = Skipthoughts_Embedding_Module( 41 | vocab= vocab_words, 42 | opt = opt 43 | ) 44 | 45 | # weight 46 | self.gw = opt['global_local_weight']['global'] 47 | self.lw = opt['global_local_weight']['local'] 48 | 49 | self.Eiters = 0 50 | 51 | self.model_name = 'AMFMN' 52 | 53 | def forward(self, img, input_local_rep, input_local_adj, text, text_lens=None): 54 | 55 | # extract features 56 | lower_feature, higher_feature, solo_feature = self.extract_feature(img) 57 | 58 | # mvsa featrues 59 | global_feature = self.mvsa(lower_feature, higher_feature, solo_feature) 60 | 61 | # text features 62 | text_feature = self.text_feature(text) 63 | 64 | sims = cosine_sim(global_feature, text_feature) 65 | return sims 66 | 67 | def factory(opt, vocab_words, cuda=True, data_parallel=True): 68 | opt = copy.copy(opt) 69 | 70 | model = BaseModel(opt, vocab_words) 71 | 72 | if data_parallel: 73 | model = nn.DataParallel(model).cuda() 74 | if not cuda: 75 | raise ValueError 76 | 77 | if cuda: 78 | model.cuda() 79 | 80 | return model 81 | 82 | 83 | 84 | -------------------------------------------------------------------------------- /predict/layers/AMFMN_utils.py: -------------------------------------------------------------------------------- 1 | # ----------------------------------------------------------- 2 | # "Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information" 3 | # Yuan, Zhiqiang and Zhang, Wenkai and Changyuan Tian and Xuee, Rong and Zhengyuan Zhang and Wang, Hongqi and Fu, Kun and Sun, Xian 4 | # Writen by YuanZhiqiang, 2021. Our code is depended on AMFMN 5 | # ------------------------------------------------------------ 6 | import math 7 | 8 | import torch 9 | import torch.nn as nn 10 | import torch.nn.functional as F 11 | import torch.nn.init 12 | from layers import seq2vec 13 | from torchvision.models.resnet import resnet18 14 | 15 | 16 | class FC(nn.Module): 17 | def __init__(self, in_size, out_size, dropout_r=0., use_relu=True): 18 | super(FC, self).__init__() 19 | self.dropout_r = dropout_r 20 | self.use_relu = use_relu 21 | 22 | self.linear = nn.Linear(in_size, out_size) 23 | 24 | if use_relu: 25 | self.relu = nn.ReLU(inplace=True) 26 | 27 | if dropout_r > 0: 28 | self.dropout = nn.Dropout(dropout_r) 29 | 30 | def forward(self, x): 31 | x = self.linear(x) 32 | 33 | if self.use_relu: 34 | x = self.relu(x) 35 | 36 | if self.dropout_r > 0: 37 | x = self.dropout(x) 38 | 39 | return x 40 | 41 | 42 | class MLP(nn.Module): 43 | def __init__(self, in_size, mid_size, out_size, dropout_r=0., use_relu=True): 44 | super(MLP, self).__init__() 45 | 46 | self.fc = FC(in_size, mid_size, dropout_r=dropout_r, use_relu=use_relu) 47 | self.linear = nn.Linear(mid_size, out_size) 48 | 49 | def forward(self, x): 50 | out = self.fc(x) 51 | return self.linear(out) 52 | 53 | 54 | class LayerNorm(nn.Module): 55 | def __init__(self, size, eps=1e-6): 56 | super(LayerNorm, self).__init__() 57 | self.eps = eps 58 | 59 | self.a_2 = nn.Parameter(torch.ones(size)) 60 | self.b_2 = nn.Parameter(torch.zeros(size)) 61 | 62 | def forward(self, x): 63 | mean = x.mean(-1, keepdim=True) 64 | std = x.std(-1, keepdim=True) 65 | 66 | return self.a_2 * (x - mean) / (std + self.eps) + self.b_2 67 | 68 | 69 | 70 | # ------------------------------ 71 | # ---- Multi-Head Attention ---- 72 | # ------------------------------ 73 | 74 | class MHAtt(nn.Module): 75 | def __init__(self, __C): 76 | super(MHAtt, self).__init__() 77 | self.__C = __C 78 | 79 | self.linear_v = nn.Linear(__C['fusion']['mca_HIDDEN_SIZE'], __C['fusion']['mca_HIDDEN_SIZE']) 80 | self.linear_k = nn.Linear(__C['fusion']['mca_HIDDEN_SIZE'], __C['fusion']['mca_HIDDEN_SIZE']) 81 | self.linear_q = nn.Linear(__C['fusion']['mca_HIDDEN_SIZE'], __C['fusion']['mca_HIDDEN_SIZE']) 82 | self.linear_merge = nn.Linear(__C['fusion']['mca_HIDDEN_SIZE'], __C['fusion']['mca_HIDDEN_SIZE']) 83 | 84 | self.dropout = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 85 | 86 | def forward(self, v, k, q, mask=None): 87 | n_batches = q.size(0) 88 | 89 | v = self.linear_v(v).view( 90 | n_batches, 91 | -1, 92 | self.__C['fusion']['mca_MULTI_HEAD'], 93 | self.__C['fusion']['mca_HIDDEN_SIZE_HEAD'] 94 | ).transpose(1, 2) 95 | 96 | k = self.linear_k(k).view( 97 | n_batches, 98 | -1, 99 | self.__C['fusion']['mca_MULTI_HEAD'], 100 | self.__C['fusion']['mca_HIDDEN_SIZE_HEAD'] 101 | ).transpose(1, 2) 102 | 103 | q = self.linear_q(q).view( 104 | n_batches, 105 | -1, 106 | self.__C['fusion']['mca_MULTI_HEAD'], 107 | self.__C['fusion']['mca_HIDDEN_SIZE_HEAD'] 108 | ).transpose(1, 2) 109 | 110 | atted = self.att(v, k, q, mask) 111 | atted = atted.transpose(1, 2).contiguous().view( 112 | n_batches, 113 | -1, 114 | self.__C['fusion']['mca_HIDDEN_SIZE'] 115 | ) 116 | 117 | atted = self.linear_merge(atted) 118 | 119 | return atted 120 | 121 | def att(self, value, key, query, mask=None): 122 | d_k = query.size(-1) 123 | 124 | scores = torch.matmul( 125 | query, key.transpose(-2, -1) 126 | ) / math.sqrt(d_k) 127 | 128 | if mask is not None: 129 | scores = scores.masked_fill(mask, -1e9) 130 | 131 | att_map = F.softmax(scores, dim=-1) 132 | att_map = self.dropout(att_map) 133 | 134 | return torch.matmul(att_map, value) 135 | 136 | 137 | # --------------------------- 138 | # ---- Feed Forward Nets ---- 139 | # --------------------------- 140 | 141 | class FFN(nn.Module): 142 | def __init__(self, __C): 143 | super(FFN, self).__init__() 144 | 145 | self.mlp = MLP( 146 | in_size=__C['fusion']['mca_HIDDEN_SIZE'], 147 | mid_size=__C['fusion']['mca_FF_SIZE'], 148 | out_size=__C['fusion']['mca_HIDDEN_SIZE'], 149 | dropout_r=__C['fusion']['mca_DROPOUT_R'], 150 | use_relu=True 151 | ) 152 | 153 | def forward(self, x): 154 | return self.mlp(x) 155 | 156 | 157 | # ------------------------ 158 | # ---- Self Attention ---- 159 | # ------------------------ 160 | 161 | class SA(nn.Module): 162 | def __init__(self, __C): 163 | super(SA, self).__init__() 164 | 165 | self.mhatt = MHAtt(__C) 166 | self.ffn = FFN(__C) 167 | 168 | self.dropout1 = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 169 | self.norm1 = LayerNorm(__C['fusion']['mca_HIDDEN_SIZE']) 170 | 171 | self.dropout2 = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 172 | self.norm2 = LayerNorm(__C['fusion']['mca_HIDDEN_SIZE']) 173 | 174 | def forward(self, x, x_mask=None): 175 | x = self.norm1(x + self.dropout1( 176 | self.mhatt(x, x, x, x_mask) 177 | )) 178 | 179 | x = self.norm2(x + self.dropout2( 180 | self.ffn(x) 181 | )) 182 | 183 | return x 184 | 185 | 186 | # ------------------------------- 187 | # ---- Self Guided Attention ---- 188 | # ------------------------------- 189 | 190 | class SGA(nn.Module): 191 | def __init__(self, __C): 192 | super(SGA, self).__init__() 193 | 194 | self.mhatt1 = MHAtt(__C) 195 | self.mhatt2 = MHAtt(__C) 196 | self.ffn = FFN(__C) 197 | 198 | self.dropout1 = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 199 | self.norm1 = LayerNorm(__C['fusion']['mca_HIDDEN_SIZE']) 200 | 201 | self.dropout2 = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 202 | self.norm2 = LayerNorm(__C['fusion']['mca_HIDDEN_SIZE']) 203 | 204 | self.dropout3 = nn.Dropout(__C['fusion']['mca_DROPOUT_R']) 205 | self.norm3 = LayerNorm(__C['fusion']['mca_HIDDEN_SIZE']) 206 | 207 | def forward(self, x, y, x_mask=None, y_mask=None): 208 | x = self.norm1(x + self.dropout1( 209 | self.mhatt1(x, x, x, x_mask) 210 | )) 211 | 212 | x = self.norm2(x + self.dropout2( 213 | self.mhatt2(y, y, x, y_mask) 214 | )) 215 | 216 | x = self.norm3(x + self.dropout3( 217 | self.ffn(x) 218 | )) 219 | 220 | return x 221 | 222 | 223 | def l2norm(X, dim, eps=1e-8): 224 | """L2-normalize columns of X 225 | """ 226 | norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps 227 | X = torch.div(X, norm) 228 | return X 229 | 230 | class ExtractFeature(nn.Module): 231 | def __init__(self, opt = {}, finetune=True): 232 | super(ExtractFeature, self).__init__() 233 | 234 | self.embed_dim = opt['embed']['embed_dim'] 235 | 236 | self.resnet = resnet18(pretrained=True) 237 | for param in self.resnet.parameters(): 238 | param.requires_grad = finetune 239 | 240 | self.pool_2x2 = nn.MaxPool2d(4) 241 | 242 | self.up_sample_2 = nn.Upsample(scale_factor=2, mode='nearest') 243 | self.up_sample_4 = nn.Upsample(scale_factor=4, mode='nearest') 244 | 245 | self.linear = nn.Linear(in_features=512, out_features=self.embed_dim) 246 | 247 | def forward(self, img): 248 | x = self.resnet.conv1(img) 249 | x = self.resnet.bn1(x) 250 | x = self.resnet.relu(x) 251 | x = self.resnet.maxpool(x) 252 | 253 | f1 = self.resnet.layer1(x) 254 | f2 = self.resnet.layer2(f1) 255 | f3 = self.resnet.layer3(f2) 256 | f4 = self.resnet.layer4(f3) 257 | 258 | # Lower Feature 259 | f2_up = self.up_sample_2(f2) 260 | lower_feature = torch.cat([f1, f2_up], dim=1) 261 | 262 | # Higher Feature 263 | f4_up = self.up_sample_2(f4) 264 | higher_feature = torch.cat([f3, f4_up], dim=1) 265 | # higher_feature = self.up_sample_4(higher_feature) 266 | 267 | # batch * 512 268 | feature = f4.view(f4.shape[0], 512, -1) 269 | solo_feature = self.linear(torch.mean(feature,dim=-1)) 270 | 271 | # torch.Size([10, 192, 64, 64]) 272 | # torch.Size([10, 768, 64, 64]) 273 | # torch.Size([10, 512]) 274 | return lower_feature, higher_feature, solo_feature 275 | 276 | class VSA_Module(nn.Module): 277 | def __init__(self, opt = {}): 278 | super(VSA_Module, self).__init__() 279 | 280 | # extract value 281 | channel_size = opt['multiscale']['multiscale_input_channel'] 282 | out_channels = opt['multiscale']['multiscale_output_channel'] 283 | embed_dim = opt['embed']['embed_dim'] 284 | 285 | # sub sample 286 | self.LF_conv = nn.Conv2d(in_channels=192, out_channels=channel_size, kernel_size=3, stride=4) 287 | self.HF_conv = nn.Conv2d(in_channels=768, out_channels=channel_size, kernel_size=1, stride=1) 288 | 289 | # visual attention 290 | self.conv1x1_1 = nn.Conv2d(in_channels=channel_size*2, out_channels=out_channels, kernel_size=1) 291 | self.conv1x1_2 = nn.Conv2d(in_channels=channel_size*2, out_channels=out_channels, kernel_size=1) 292 | 293 | # solo attention 294 | self.solo_attention = nn.Linear(in_features=256, out_features=embed_dim) 295 | 296 | def forward(self, lower_feature, higher_feature, solo_feature): 297 | 298 | # b x channel_size x 16 x 16 299 | lower_feature = self.LF_conv(lower_feature) 300 | higher_feature = self.HF_conv(higher_feature) 301 | 302 | # concat 303 | concat_feature = torch.cat([lower_feature, higher_feature], dim=1) 304 | 305 | # residual 306 | concat_feature = higher_feature.mean(dim=1,keepdim=True).expand_as(concat_feature) + concat_feature 307 | 308 | # attention 309 | main_feature = self.conv1x1_1(concat_feature) 310 | attn_feature = torch.sigmoid(self.conv1x1_2(concat_feature).view(concat_feature.shape[0],1,-1)).view(concat_feature.shape[0], 1, main_feature.shape[2], main_feature.shape[3]) 311 | atted_feature = (main_feature*attn_feature).squeeze(dim=1).view(attn_feature.shape[0], -1) 312 | 313 | # solo attention 314 | solo_att = torch.sigmoid(self.solo_attention(atted_feature)) 315 | solo_feature = solo_feature*solo_att 316 | 317 | return l2norm(solo_feature, -1) 318 | 319 | class Skipthoughts_Embedding_Module(nn.Module): 320 | def __init__(self, vocab, opt, out_dropout=-1): 321 | super(Skipthoughts_Embedding_Module, self).__init__() 322 | self.opt = opt 323 | self.vocab_words = vocab 324 | 325 | self.seq2vec = seq2vec.factory(self.vocab_words, self.opt['seq2vec'], self.opt['seq2vec']['dropout']) 326 | 327 | self.to_out = nn.Linear(in_features=2400, out_features=self.opt['embed']['embed_dim']) 328 | self.dropout = out_dropout 329 | 330 | def forward(self, input_text ): 331 | x_t_vec = self.seq2vec(input_text) 332 | out = F.relu(self.to_out(x_t_vec)) 333 | if self.dropout >= 0: 334 | out = F.dropout(out, self.dropout) 335 | 336 | return out 337 | 338 | def params_count(model): 339 | count = 0 340 | for p in model.parameters(): 341 | c = 1 342 | for i in range(p.dim()): 343 | c *= p.size(i) 344 | count += c 345 | return count 346 | 347 | def bcosine_sim(im, s): 348 | """Cosine similarity between all the image and sentence pairs 349 | """ 350 | im = l2norm(im, dim=-1) 351 | s = l2norm(s, dim=-1) 352 | w12 = im.mm(s.t()) 353 | return w12 354 | # ==================================================================== 355 | # About GCN 356 | class GCN(nn.Module): 357 | def __init__(self , dim_in=20 , dim_out=20, dim_embed = 512): 358 | super(GCN,self).__init__() 359 | 360 | self.fc1 = nn.Linear(dim_in ,dim_in,bias=False) 361 | self.fc2 = nn.Linear(dim_in,dim_in//2,bias=False) 362 | self.fc3 = nn.Linear(dim_in//2,dim_out,bias=False) 363 | 364 | self.out = nn.Linear(dim_out * dim_in, dim_embed) 365 | 366 | def forward(self, A, X): 367 | batch, objects, rep = X.shape[0], X.shape[1], X.shape[2] 368 | 369 | # first layer 370 | tmp = (A.bmm(X)).view(-1, rep) 371 | X = F.relu(self.fc1(tmp)) 372 | X = X.view(batch, -1, X.shape[-1]) 373 | 374 | # second layer 375 | tmp = (A.bmm(X)).view(-1, X.shape[-1]) 376 | X = F.relu(self.fc2(tmp)) 377 | X = X.view(batch, -1, X.shape[-1]) 378 | 379 | # third layer 380 | tmp = (A.bmm(X)).view(-1, X.shape[-1]) 381 | X = F.relu(self.fc3(tmp)) 382 | X = X.view(batch, -1) 383 | 384 | return l2norm(self.out(X), -1) 385 | 386 | 387 | 388 | 389 | 390 | 391 | 392 | 393 | 394 | 395 | 396 | 397 | 398 | 399 | 400 | 401 | 402 | -------------------------------------------------------------------------------- /predict/layers/LW_MCR.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/05 6 | 7 | import copy 8 | 9 | from .MCR_utils import * 10 | 11 | 12 | class unsupervised_Visual_Model(nn.Module): 13 | def __init__(self, feature_dim=128): 14 | super(unsupervised_Visual_Model, self).__init__() 15 | 16 | self.extract_feature = ExtractFeature() 17 | 18 | # projection head 19 | self.g = nn.Sequential(nn.Linear(512, 512, bias=False), nn.BatchNorm1d(512), 20 | nn.ReLU(inplace=True), nn.Linear(512, feature_dim, bias=True)) 21 | 22 | def forward(self, x): 23 | feature, _, _ = self.extract_feature(x) 24 | out = self.g(feature) 25 | return F.normalize(feature, dim=-1), F.normalize(out, dim=-1) 26 | 27 | class unsupervised_Text_Model(nn.Module): 28 | def __init__(self, feature_dim=128, vocab_words=[]): 29 | super(unsupervised_Text_Model, self).__init__() 30 | 31 | 32 | self.text_feature = textCNN( 33 | vocab= vocab_words, 34 | opt = None 35 | ) 36 | # projection head 37 | self.g = nn.Sequential(nn.Linear(512, 512, bias=False), nn.BatchNorm1d(512), 38 | nn.ReLU(inplace=True), nn.Linear(512, feature_dim, bias=True)) 39 | 40 | def forward(self, x): 41 | feature = self.text_feature(x) 42 | out = self.g(feature) 43 | return F.normalize(feature, dim=-1), F.normalize(out, dim=-1) 44 | 45 | class BaseModel(nn.Module): 46 | def __init__(self, opt={}, vocab_words=[]): 47 | super(BaseModel, self).__init__() 48 | 49 | # img feature 50 | self.extract_feature = ExtractFeature() 51 | 52 | # self.pre_train_extroctor = unsupervised_Visual_Model() 53 | # state = torch.load('./model/unsupervised_pretrain_model/visual_pre.pth') 54 | # self.pre_train_extroctor.load_state_dict(state) 55 | # del state 56 | # self.extract_feature = self.pre_train_extroctor.extract_feature 57 | # del self.pre_train_extroctor 58 | 59 | 60 | 61 | self.text_feature = textCNN( 62 | vocab= vocab_words, 63 | opt = opt 64 | ) 65 | 66 | # self.pre_train_extroctor = unsupervised_Text_Model(vocab_words= vocab_words,) 67 | # state = torch.load('./model/unsupervised_pretrain_model/text_pre.pth') 68 | # self.pre_train_extroctor.load_state_dict(state) 69 | # del state 70 | # self.text_feature = self.pre_train_extroctor.text_feature 71 | # del self.pre_train_extroctor 72 | 73 | self.Eiters = 0 74 | 75 | self.model_name = 'LW_MCR' 76 | 77 | 78 | def forward(self, img, text, text_lens=None): 79 | 80 | # extract features 81 | visual_feature, lower_feature, higher_feature = self.extract_feature(img) 82 | 83 | # text features 84 | text_feature = self.text_feature(text) 85 | 86 | # print("visual_feature.shape:{}".format(visual_feature.shape)) 87 | # print("text_feature.shape:{}".format(text_feature.shape)) 88 | # visual_feature.shape: torch.Size([8, 512]) 89 | # text_feature.shape: torch.Size([8, 512]) 90 | # print("=========================") 91 | # exit(0) 92 | 93 | sims = cosine_sim(visual_feature, text_feature) 94 | 95 | return sims, [visual_feature, text_feature, lower_feature, higher_feature] 96 | 97 | def factory(opt, vocab_words, cuda=True, data_parallel=True,device_ids=[0]): 98 | opt = copy.copy(opt) 99 | 100 | model = BaseModel(opt, vocab_words) 101 | 102 | if data_parallel: 103 | model = nn.DataParallel(model, device_ids=device_ids).cuda() 104 | if not cuda: 105 | raise ValueError 106 | 107 | if cuda: 108 | model.cuda() 109 | 110 | return model 111 | 112 | 113 | -------------------------------------------------------------------------------- /predict/layers/MCR_utils.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/05 6 | 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | import torch.nn.init 11 | 12 | 13 | def l2norm(X, dim, eps=1e-8): 14 | """L2-normalize columns of X 15 | """ 16 | norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps 17 | X = torch.div(X, norm) 18 | return X 19 | 20 | # ---------------------------------------------------------------------------- 21 | # ------------------------- Visual Feature Encoding ------------------ 22 | 23 | class ChannelAttention(nn.Module): 24 | def __init__(self, in_planes, ratio=16): 25 | super(ChannelAttention, self).__init__() 26 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 27 | self.max_pool = nn.AdaptiveMaxPool2d(1) 28 | 29 | self.fc = nn.Sequential(nn.Conv2d(in_planes, in_planes // 4, 1, bias=False), 30 | nn.ReLU(), 31 | nn.Conv2d(in_planes // 4, in_planes, 1, bias=False)) 32 | self.sigmoid = nn.Sigmoid() 33 | 34 | def forward(self, x): 35 | avg_out = self.fc(self.avg_pool(x)) 36 | max_out = self.fc(self.max_pool(x)) 37 | out = avg_out + max_out 38 | return self.sigmoid(out) 39 | 40 | 41 | class SpatialAttention(nn.Module): 42 | def __init__(self, kernel_size=7): 43 | super(SpatialAttention, self).__init__() 44 | 45 | self.conv1 = nn.Conv2d(2, 1, kernel_size, padding=kernel_size // 2, bias=False) 46 | self.sigmoid = nn.Sigmoid() 47 | 48 | def forward(self, x): 49 | avg_out = torch.mean(x, dim=1, keepdim=True) 50 | max_out, _ = torch.max(x, dim=1, keepdim=True) 51 | x = torch.cat([avg_out, max_out], dim=1) 52 | x = self.conv1(x) 53 | return self.sigmoid(x) 54 | 55 | class Fire(nn.Module): 56 | 57 | def __init__(self, inplanes, squeeze_planes, 58 | expand1x1_planes, expand3x3_planes): 59 | super(Fire, self).__init__() 60 | self.inplanes = inplanes 61 | self.squeeze = nn.Conv2d(inplanes, squeeze_planes, kernel_size=1) 62 | self.squeeze_activation = nn.ReLU(inplace=True) 63 | self.expand1x1 = nn.Conv2d(squeeze_planes, expand1x1_planes, 64 | kernel_size=1) 65 | self.expand1x1_activation = nn.ReLU(inplace=True) 66 | self.expand3x3 = nn.Conv2d(squeeze_planes, expand3x3_planes, 67 | kernel_size=3, padding=1) 68 | self.expand3x3_activation = nn.ReLU(inplace=True) 69 | 70 | def forward(self, x): 71 | x = self.squeeze_activation(self.squeeze(x)) 72 | return torch.cat([ 73 | self.expand1x1_activation(self.expand1x1(x)), 74 | self.expand3x3_activation(self.expand3x3(x)) 75 | ], 1) 76 | 77 | class redefine_squeezenet(nn.Module): 78 | def __init__(self, ): 79 | super(redefine_squeezenet, self).__init__() 80 | 81 | self.conv_1 = nn.Sequential( 82 | nn.Conv2d(3, 64, kernel_size=3, stride=2), 83 | nn.ReLU(inplace=True), 84 | nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), 85 | Fire(64, 16, 64, 64), 86 | Fire(128, 16, 64, 64), 87 | ) 88 | self.conv_2 = nn.Sequential( 89 | nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), 90 | Fire(128, 32, 128, 128), 91 | Fire(256, 32, 128, 128), 92 | nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), 93 | Fire(256, 48, 192, 192), 94 | Fire(384, 48, 192, 192), 95 | Fire(384, 64, 256, 256), 96 | Fire(512, 64, 256, 256), 97 | ) 98 | # Final convolution is initialized differently from the rest 99 | final_conv = nn.Conv2d(512, 1000, kernel_size=1) 100 | self.classifier = nn.Sequential( 101 | nn.Dropout(p=0.5), 102 | final_conv, 103 | nn.ReLU(inplace=True), 104 | nn.AdaptiveAvgPool2d((1, 1)) 105 | ) 106 | 107 | def forward(self, img): 108 | low_feature = self.conv_1(img) 109 | high_feature = self.conv_2(low_feature) 110 | results = self.classifier(high_feature) 111 | return results 112 | 113 | def convert_squeezenet(squeeze, net):#vgg16是pytorch自带的 114 | vgg_items = net.state_dict().items() 115 | vgg16_items = squeeze.items() 116 | 117 | pretrain_model = {} 118 | j = 0 119 | for k, v in net.state_dict().items():#按顺序依次填入 120 | v = list(vgg16_items)[j][1] 121 | k = list(vgg_items)[j][0] 122 | pretrain_model[k] = v 123 | j += 1 124 | return pretrain_model 125 | 126 | class extract_by_squeezenet(nn.Module): 127 | def __init__(self, ): 128 | super(extract_by_squeezenet, self).__init__() 129 | 130 | net = redefine_squeezenet() 131 | 132 | self.conv_1 = net.conv_1 133 | self.conv_2 = net.conv_2 134 | 135 | def forward(self, img): 136 | low_feature = self.conv_1(img) 137 | high_feature = self.conv_2(low_feature) 138 | return low_feature, high_feature 139 | 140 | class Pretrain_visual_extractor(nn.Module): 141 | def __init__(self, parms=None): 142 | super(Pretrain_visual_extractor, self).__init__() 143 | # backbone = resnet18(pretrained=True) 144 | # self.backbone = myResnet(backbone) 145 | self.backbone = extract_by_squeezenet() 146 | 147 | # lower and higher transformation 148 | self.lower_trans = nn.Sequential( 149 | nn.Conv2d(in_channels=128, out_channels=64, stride=2, kernel_size=3, padding=1), 150 | nn.BatchNorm2d(num_features=64), 151 | nn.ReLU(inplace=True) 152 | ) 153 | self.higher_trans = nn.Sequential( 154 | nn.Upsample(scale_factor=2, mode='nearest'), 155 | nn.Conv2d(in_channels=512, out_channels=64, stride=1, kernel_size=1, padding=1), 156 | nn.BatchNorm2d(num_features=64), 157 | nn.ReLU(inplace=True) 158 | ) 159 | 160 | # BCNN 161 | 162 | # image self attention 163 | self.ISA_1_c = nn.Conv2d(in_channels=1, out_channels=4, kernel_size=3, stride=2, padding=1) 164 | self.ISA_1_b = nn.BatchNorm2d(4) 165 | self.ISA_1_p = nn.PReLU() 166 | self.ISA_1_ca = ChannelAttention(4) 167 | self.ISA_1_sa = SpatialAttention() 168 | self.ISA_1_pool = nn.MaxPool2d(kernel_size=2) 169 | 170 | self.ISA_2_c = nn.Conv2d(in_channels=4, out_channels=8, kernel_size=3, stride=2, padding=1) 171 | self.ISA_2_b = nn.BatchNorm2d(8) 172 | self.ISA_2_p = nn.PReLU() 173 | self.ISA_2_ca = ChannelAttention(8) 174 | self.ISA_2_sa = SpatialAttention() 175 | 176 | self.fc = torch.nn.Linear(64*8, 10) 177 | 178 | def forward(self, img): 179 | N = img.size()[0] 180 | 181 | # backbone 182 | lower_feature, higher_feature = self.backbone(img) 183 | 184 | # lower and higher transformation 185 | lower_feature = self.lower_trans(lower_feature) 186 | higher_feature = self.higher_trans(higher_feature) 187 | 188 | # BCNN 189 | lower_feature = lower_feature.view(N, 64, 32*32) 190 | higher_feature = higher_feature.view(N, 64, 32*32) 191 | X = torch.bmm(lower_feature, torch.transpose(higher_feature, 1, 2)) / (32 ** 2) 192 | 193 | X = torch.sqrt(X + 1e-5) 194 | X = torch.nn.functional.normalize(X) 195 | X = torch.unsqueeze(X, dim=1) 196 | 197 | # image self attention 198 | X = self.ISA_1_c(X) 199 | X = self.ISA_1_b(X) 200 | X = self.ISA_1_p(X) 201 | X = self.ISA_1_ca(X) * X 202 | X = self.ISA_1_sa(X) * X 203 | X = self.ISA_1_pool(X) 204 | 205 | X = self.ISA_2_c(X) 206 | X = self.ISA_2_b(X) 207 | X = self.ISA_2_p(X) 208 | X = self.ISA_2_ca(X) * X 209 | X = self.ISA_2_sa(X) * X 210 | 211 | X = X.view(N, -1) 212 | X = self.fc(X) 213 | 214 | return X 215 | 216 | class ExtractFeature(nn.Module): 217 | def __init__(self): 218 | super(ExtractFeature, self).__init__() 219 | self.pre_train_extroctor = Pretrain_visual_extractor() 220 | 221 | self.backbone = self.pre_train_extroctor.backbone 222 | 223 | # lower and higher transformation 224 | self.lower_trans = self.pre_train_extroctor.lower_trans 225 | self.higher_trans = self.pre_train_extroctor.higher_trans 226 | 227 | # BCNN 228 | 229 | # image self attention 230 | self.ISA_1_c = self.pre_train_extroctor.ISA_1_c 231 | self.ISA_1_b = self.pre_train_extroctor.ISA_1_b 232 | self.ISA_1_p = self.pre_train_extroctor.ISA_1_p 233 | self.ISA_1_ca = self.pre_train_extroctor.ISA_1_ca 234 | self.ISA_1_sa = self.pre_train_extroctor.ISA_1_sa 235 | self.ISA_1_pool = self.pre_train_extroctor.ISA_1_pool 236 | self.dropout_1 = nn.Dropout(0.2) 237 | 238 | self.ISA_2_c = self.pre_train_extroctor.ISA_2_c 239 | self.ISA_2_b = self.pre_train_extroctor.ISA_2_b 240 | self.ISA_2_p = self.pre_train_extroctor.ISA_2_p 241 | self.ISA_2_ca = self.pre_train_extroctor.ISA_2_ca 242 | self.ISA_2_sa = self.pre_train_extroctor.ISA_2_sa 243 | self.dropout_2 = nn.Dropout(0.2) 244 | 245 | 246 | del self.pre_train_extroctor 247 | 248 | self.fc = nn.Linear(512, 512) 249 | 250 | def forward(self, img): 251 | 252 | N = img.size()[0] 253 | 254 | # backbone 255 | lower_feature, higher_feature = self.backbone(img) 256 | 257 | # lower and higher transformation 258 | lf = self.lower_trans(lower_feature) 259 | hf = self.higher_trans(higher_feature) 260 | # print("lower_feature.shape:{}".format(lower_feature.shape)) 261 | # print("higher_feature.shape:{}".format(higher_feature.shape)) 262 | # lower_feature.shape: torch.Size([8, 64, 32, 32]) 263 | # higher_feature.shape: torch.Size([8, 64, 32, 32]) 264 | # print("=========================") 265 | 266 | # BCNN 267 | lower_feature = lf.view(N, 64, 32*32) 268 | higher_feature = hf.view(N, 64, 32*32) 269 | X = torch.bmm(lower_feature, torch.transpose(higher_feature, 1, 2)) / (32 ** 2) 270 | 271 | X = torch.sqrt(X + 1e-5) 272 | X = torch.nn.functional.normalize(X) 273 | X = torch.unsqueeze(X, dim=1) 274 | 275 | # print("fusion.shape:{}".format(X.shape)) 276 | # fusion.shape: torch.Size([8, 1, 64, 64]) 277 | # print("=========================") 278 | 279 | # image self attention 280 | X = self.ISA_1_c(X) 281 | X = self.ISA_1_b(X) 282 | X = self.ISA_1_p(X) 283 | X = self.ISA_1_ca(X) * X 284 | X = self.ISA_1_sa(X) * X 285 | X = self.ISA_1_pool(X) 286 | X = self.dropout_1(X) 287 | 288 | X = self.ISA_2_c(X) 289 | X = self.ISA_2_b(X) 290 | X = self.ISA_2_p(X) 291 | X = self.ISA_2_ca(X) * X 292 | X = self.ISA_2_sa(X) * X 293 | X = self.dropout_2(X) 294 | 295 | X = X.view(N, -1) 296 | 297 | X = self.fc(X) 298 | 299 | return l2norm(X, -1), lf, hf 300 | 301 | def params_count(model): 302 | count = 0 303 | for p in model.parameters(): 304 | c = 1 305 | for i in range(p.dim()): 306 | c *= p.size(i) 307 | count += c 308 | return count 309 | # 310 | # model = ExtractFeature() 311 | # from torch.autograd import Variable 312 | # input = Variable(torch.zeros(10, 3, 256, 256)) 313 | # feature = model(input) 314 | # print(feature.shape) 315 | # print(params_count(model)) 316 | # 317 | # exit() 318 | 319 | # ---------------------------------------------------------------------------- 320 | # ------------------------- Text Feature Encoding ------------------ 321 | 322 | class textCNN(nn.Module): 323 | def __init__(self, vocab, opt, lstm_dropout=0.25, out_dropout=-1): 324 | super(textCNN, self).__init__() 325 | Vocab = len(vocab)+1 ## 已知词的数量 326 | Dim = 300 ##每个词向量长度 327 | Ci = 1 ##输入的channel数 328 | Knum = 100 ## 每种卷积核的数量 329 | Ks = [3,4,5] ## 卷积核list，形如[2,3,4] 330 | Cla = 512 331 | 332 | self.embed = nn.Embedding(Vocab, Dim) ## 词向量，这里直接随机 333 | self.convs = nn.ModuleList([nn.Conv2d(Ci, Knum, (K, Dim)) for K in Ks]) ## 卷积层 334 | 335 | self.dropout = nn.Dropout(lstm_dropout) 336 | self.fc = nn.Linear(len(Ks)*Knum,Cla) ##全连接层 337 | 338 | def forward(self, x): 339 | x = self.embed(x) # (8, 18, 300) 340 | 341 | x = x.unsqueeze(1) # (8, 1, 18, 300) 342 | 343 | x = [F.relu(conv(x)).squeeze(3) for conv in self.convs] # len(Ks)*(N,Knum,W) 344 | 345 | x = [F.max_pool1d(line, line.size(2)).squeeze(2) for line in x] # len(Ks)*(N,Knum) 346 | 347 | x = torch.cat(x, 1) # (N,Knum*len(Ks)) 348 | 349 | x = self.dropout(x) # (N, len(Ks)*Co) 350 | logit = self.fc(x) # (N, C) 351 | 352 | return l2norm(logit, -1) 353 | 354 | def cosine_sim(im, s): 355 | """Cosine similarity between all the image and sentence pairs 356 | """ 357 | # w1 = l2norm(im, dim=-1) 358 | # w2 = l2norm(s, dim=-1) 359 | w12 = im.mm(s.t()) 360 | return w12 361 | -------------------------------------------------------------------------------- /predict/layers/__pycache__/AMFMN.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/layers/__pycache__/AMFMN.cpython-37.pyc -------------------------------------------------------------------------------- /predict/layers/__pycache__/AMFMN_utils.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/layers/__pycache__/AMFMN_utils.cpython-37.pyc -------------------------------------------------------------------------------- /predict/layers/__pycache__/seq2vec.cpython-37.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/layers/__pycache__/seq2vec.cpython-37.pyc -------------------------------------------------------------------------------- /predict/layers/seq2vec.py: -------------------------------------------------------------------------------- 1 | # ----------------------------------------------------------- 2 | # "Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information" 3 | # Yuan, Zhiqiang and Zhang, Wenkai and Changyuan Tian and Xuee, Rong and Zhengyuan Zhang and Wang, Hongqi and Fu, Kun and Sun, Xian 4 | # Writen by YuanZhiqiang, 2021. Our code is depended on AMFMN 5 | # ------------------------------------------------------------ 6 | 7 | # A revision version from Skip-thoughs 8 | import skipthoughts 9 | 10 | 11 | def factory(vocab_words, opt , dropout=0.25): 12 | if opt['arch'] == 'skipthoughts': 13 | st_class = getattr(skipthoughts, opt['type']) 14 | seq2vec = st_class(opt['dir_st'], 15 | vocab_words, 16 | dropout=dropout, 17 | fixed_emb=opt['fixed_emb']) 18 | 19 | else: 20 | raise NotImplementedError 21 | return seq2vec 22 | -------------------------------------------------------------------------------- /predict/model_encoder.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/05 6 | 7 | import nltk 8 | import numpy as np 9 | import torch 10 | import torchvision.transforms as transforms 11 | from PIL import Image 12 | 13 | 14 | def l2norm(X, dim, eps=1e-8): 15 | """L2-normalize columns of X 16 | """ 17 | norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps 18 | X = torch.div(X, norm) 19 | return X 20 | 21 | # Encoder for LW_MCR 22 | class EncoderLWMCR: 23 | def image_encoder(self, model, image_path): 24 | transform = transforms.Compose([ 25 | transforms.Resize((256, 256)), 26 | transforms.ToTensor(), 27 | transforms.Normalize((0.485, 0.456, 0.406), 28 | (0.229, 0.224, 0.225))]) 29 | 30 | # data preprocessing 31 | image = Image.open(image_path).convert('RGB') 32 | image = transform(image) # torch.Size([3, 256, 256]) 33 | image = torch.unsqueeze(image, dim=0).cuda() 34 | 35 | visual_feature, lower_feature, higher_feature = model.extract_feature(image) 36 | global_feature = l2norm(visual_feature, dim=-1) 37 | 38 | # to cpu vector 39 | vector = global_feature.cpu().detach().numpy()[0] 40 | 41 | return vector 42 | 43 | def text_encoder(self, model, vocab, text): 44 | 45 | # Convert caption (string) to word ids. 46 | tokens = nltk.tokenize.word_tokenize(text.lower()) 47 | punctuations = [',', '.', ':', ';', '?', '(', ')', '[', ']', '&', '!', '*', '@', '#', '$', '%'] 48 | tokens = [k for k in tokens if k not in punctuations] 49 | tokens_UNK = [k if k in vocab.word2idx.keys() else '' for k in tokens] 50 | 51 | caption = [] 52 | caption.extend([vocab(token) for token in tokens_UNK]) 53 | caption = torch.LongTensor(caption) 54 | caption = torch.unsqueeze(caption, dim=0).cuda() 55 | 56 | # model processing 57 | text_feature = model.text_feature(caption) 58 | text_feature = l2norm(text_feature, dim=-1) 59 | 60 | # to cpu vector 61 | vector = text_feature.cpu().detach().numpy()[0] 62 | 63 | return vector 64 | 65 | # Encoder for AMFMN 66 | class EncoderAMFMN: 67 | def image_encoder(self, model, image_path): 68 | transform = transforms.Compose([ 69 | transforms.Resize((256, 256)), 70 | transforms.ToTensor(), 71 | transforms.Normalize((0.485, 0.456, 0.406), 72 | (0.229, 0.224, 0.225))]) 73 | # check image size 74 | # image_path = trans_bigimage_to_small(image_path) 75 | 76 | # data preprocessing 77 | image = Image.open(image_path).convert('RGB') 78 | image = transform(image) # torch.Size([3, 256, 256]) 79 | image = torch.unsqueeze(image, dim=0).cuda() 80 | 81 | # model processing 82 | lower_feature, higher_feature, solo_feature = model.extract_feature(image) 83 | global_feature = model.mvsa(lower_feature, higher_feature, solo_feature) 84 | global_feature = l2norm(global_feature, dim=-1) 85 | 86 | # to cpu vector 87 | vector = global_feature.cpu().detach().numpy()[0] 88 | 89 | return vector 90 | 91 | def text_encoder(self, model, vocab, text): 92 | 93 | # Convert caption (string) to word ids. 94 | tokens = nltk.tokenize.word_tokenize(text.lower()) 95 | punctuations = [',', '.', ':', ';', '?', '(', ')', '[', ']', '&', '!', '*', '@', '#', '$', '%'] 96 | tokens = [k for k in tokens if k not in punctuations] 97 | tokens_UNK = [k if k in vocab.word2idx.keys() else '' for k in tokens] 98 | 99 | caption = [] 100 | caption.extend([vocab(token) for token in tokens_UNK]) 101 | caption = torch.LongTensor(caption) 102 | caption = torch.unsqueeze(caption, dim=0).cuda() 103 | caption = caption.expand((2, caption.shape[-1])) 104 | 105 | # model processing 106 | text_feature = model.text_feature(caption) 107 | text_feature = l2norm(text_feature, dim=-1) 108 | 109 | # to cpu vector 110 | vector = text_feature.cpu().detach().numpy()[0] 111 | 112 | return vector 113 | 114 | class Encoder: 115 | def __init__(self, model): 116 | if model.model_name == 'AMFMN': 117 | self.encoder = EncoderAMFMN() 118 | elif model.model_name == 'LW_MCR': 119 | self.encoder = EncoderLWMCR() 120 | else: 121 | raise NotImplementedError 122 | 123 | def cosine_sim(self, image_vector, text_vector): 124 | """ 125 | 计算两个向量间的余弦相似度 126 | :param image_vector: 图片编码向量 127 | :param text_vector: 文本编码向量 128 | :return: 相似度 129 | """ 130 | if hasattr(self.encoder, 'calc_similarity'): 131 | return self.encoder.calc_similarity(image_vector, text_vector) 132 | else: 133 | image_vector = image_vector / np.linalg.norm(image_vector) 134 | text_vector = text_vector / np.linalg.norm(text_vector) 135 | 136 | similarity = np.mean(np.multiply(image_vector, text_vector)) 137 | return similarity 138 | 139 | def image_encoder(self, model, image_path): 140 | """ 141 | 提供的图像编码函数 142 | :param model: 模型文件 143 | :param image_path: 图像路径 144 | :return: 编码向量 145 | """ 146 | 147 | return self.encoder.image_encoder(model, image_path) 148 | 149 | def text_encoder(self, model, vocab, text): 150 | """ 151 | 提供的文本编码函数 152 | :param model: 模型文件 153 | :param vocab: 文本字典 154 | :param text: 编码文本 155 | :return: 编码向量 156 | """ 157 | 158 | return self.encoder.text_encoder(model, vocab, text) 159 | 160 | 161 | if __name__ == "__main__": 162 | from model_init import model_init 163 | 164 | prefix = "./" 165 | yaml_path = "option/RSITMD/RSITMD_AMFMN.yaml" 166 | test_jpg = "./test_data/sparseresidential_3814.tif" 167 | test_caption = "many airplane parked in the airport" 168 | 169 | # init model 170 | model, vocab = model_init( 171 | prefix_path = "./", 172 | yaml_path = yaml_path 173 | ) 174 | 175 | # encoder 176 | encoder = Encoder(model) 177 | visual_vector = encoder.image_encoder(model, test_jpg) 178 | text_vector = encoder.text_encoder(model, vocab, test_caption) 179 | 180 | print("visual_vector:", np.shape(visual_vector)) 181 | print("text_vector:", np.shape(text_vector)) 182 | 183 | if len(visual_vector) == len(text_vector): 184 | print("Encoder test successful!") 185 | 186 | sims = encoder.cosine_sim(visual_vector, text_vector) 187 | print("Calc sim successful!") 188 | 189 | 190 | 191 | 192 | 193 | 194 | -------------------------------------------------------------------------------- /predict/model_init.py: -------------------------------------------------------------------------------- 1 | #encoding:utf-8 2 | # ----------------------------------------------------------- 3 | # "Exploring a Fine-Grained Multiscale Method for Cross-Modal Remote Sensing Image Retrieval" 4 | # Yuan, Zhiqiang and Zhang, Wenkai and Fu, Kun and Li, Xuan and Deng, Chubo and Wang, Hongqi and Sun, Xian 5 | # IEEE Transactions on Geoscience and Remote Sensing 2021 6 | # Writen by YuanZhiqiang, 2021. Our code is depended on MTFN 7 | # ------------------------------------------------------------ 8 | 9 | import os 10 | 11 | import torch 12 | import yaml 13 | from vocabs import deserialize_vocab 14 | 15 | 16 | def parser_options(prefix_path, yaml_path): 17 | # load model options 18 | with open(os.path.join(prefix_path,yaml_path), 'r') as handle: 19 | options = yaml.safe_load(handle) 20 | 21 | return options 22 | 23 | def model_init(prefix_path, yaml_path): 24 | options = parser_options(prefix_path, yaml_path) 25 | 26 | # choose model 27 | if options['model']['name'] == "AMFMN": 28 | from layers import AMFMN as models 29 | elif options['model']['name'] == "LW_MCR": 30 | from layers import LW_MCR as models 31 | else: 32 | raise NotImplementedError 33 | 34 | # make vocab 35 | vocab = deserialize_vocab(os.path.join(prefix_path,options['dataset']['vocab_path'])) 36 | vocab_word = sorted(vocab.word2idx.items(), key=lambda x: x[1], reverse=False) 37 | vocab_word = [tup[0] for tup in vocab_word] 38 | 39 | model = models.factory(options['model'], 40 | vocab_word, 41 | cuda=True, 42 | data_parallel=False) 43 | 44 | checkpoint = torch.load(options['optim']['resume']) 45 | model.load_state_dict(checkpoint['model'], strict=True) 46 | model.eval() 47 | 48 | return model, vocab 49 | 50 | if __name__ == "__main__": 51 | prefix = "./" 52 | yaml_path = "option/RSITMD/RSITMD_LW_MCR.yaml" 53 | 54 | model, vocab = model_init( 55 | prefix_path = "./", 56 | yaml_path = yaml_path 57 | ) 58 | 59 | try: 60 | model.eval() 61 | print(vocab) 62 | print("Successfully load the model.") 63 | except Exception as e: 64 | print("Failed to load the model.") -------------------------------------------------------------------------------- /predict/option/RSITMD/RSITMD_AMFMN.yaml: -------------------------------------------------------------------------------- 1 | model: 2 | multiscale: 3 | multiscale_input_channel: 3 4 | multiscale_output_channel: 1 5 | cross_attention: 6 | att_type: "soft_att" 7 | seq2vec: 8 | arch: skipthoughts 9 | dir_st: /data 10 | type: BayesianUniSkip 11 | dropout: 0.25 12 | fixed_emb: False 13 | fusion: 14 | correct_local_hidden_dim: 512 15 | correct_local_hidden_drop: 0.2 16 | supplement_global_hidden_dim: 512 17 | supplement_global_hidden_drop: 0.2 18 | dynamic_fusion_dim: 512 19 | dynamic_fusion_drop: 0.2 20 | mca_DROPOUT_R: 0.1 21 | mca_HIDDEN_SIZE: 512 22 | mca_FF_SIZE: 1024 23 | mca_MULTI_HEAD: 8 24 | mca_HIDDEN_SIZE_HEAD: 64 25 | embed: 26 | embed_dim: 512 27 | global_local_weight: 28 | global: None 29 | local: None 30 | name: AMFMN 31 | dataset: 32 | datatype: rsitmd 33 | data_split: 34 | data_path: 'data/rsitmd_precomp/' 35 | image_path: '/data/rsitmd_images/' 36 | vocab_path: 'vocab/rsitmd_splits_vocab.json' 37 | local_path: './detection/representation/RSITMD/rsitmd_local.npy' 38 | batch_size: 100 39 | batch_size_val: 70 40 | workers: 3 41 | optim: 42 | epochs: 50 43 | lr: 0.0002 44 | lr_decay_param: 0.7 45 | lr_update_epoch: 20 46 | grad_clip: 0 47 | max_violation: 0 48 | margin: 0.2 49 | resume: "./checkpoints/AMFMN.tar" 50 | logs: 51 | eval_step: 1 52 | print_freq: 10 53 | ckpt_save_path: "checkpoints/RSITMD_AMFMN.tar" 54 | logger_name: 'logs/' 55 | k_fold: 56 | experiment_name: 'rsitmd_AMFMN' 57 | nums: 1 58 | current_num: 0 59 | -------------------------------------------------------------------------------- /predict/option/RSITMD/RSITMD_LW_MCR.yaml: -------------------------------------------------------------------------------- 1 | model: 2 | multiscale: 3 | multiscale_input_channel: 3 4 | multiscale_output_channel: 1 5 | cross_attention: 6 | att_type: "soft_att" 7 | seq2vec: 8 | arch: skipthoughts 9 | dir_st: /data 10 | type: BayesianUniSkip 11 | dropout: 0.25 12 | fixed_emb: False 13 | fusion: 14 | correct_local_hidden_dim: 512 15 | correct_local_hidden_drop: 0.2 16 | supplement_global_hidden_dim: 512 17 | supplement_global_hidden_drop: 0.2 18 | dynamic_fusion_dim: 512 19 | dynamic_fusion_drop: 0.2 20 | mca_DROPOUT_R: 0.1 21 | mca_HIDDEN_SIZE: 512 22 | mca_FF_SIZE: 1024 23 | mca_MULTI_HEAD: 8 24 | mca_HIDDEN_SIZE_HEAD: 64 25 | embed: 26 | embed_dim: 512 27 | global_local_weight: 28 | global: None 29 | local: None 30 | name: LW_MCR 31 | dataset: 32 | datatype: rsitmd 33 | data_split: 34 | data_path: 'data/rsitmd_precomp/' 35 | image_path: '/data/rsitmd_images/' 36 | vocab_path: 'vocab/rsitmd_splits_vocab.json' 37 | local_path: './detection/representation/RSITMD/rsitmd_local.npy' 38 | batch_size: 100 39 | batch_size_val: 70 40 | workers: 3 41 | optim: 42 | epochs: 50 43 | lr: 0.0002 44 | lr_decay_param: 0.7 45 | lr_update_epoch: 20 46 | grad_clip: 0 47 | max_violation: 0 48 | margin: 0.2 49 | resume: "/workspace/slm/predict/checkpoints/RSITMD_LW_MCR.tar" 50 | logs: 51 | eval_step: 1 52 | print_freq: 10 53 | ckpt_save_path: "checkpoints/LW_MCR.tar" 54 | logger_name: 'logs/' 55 | k_fold: 56 | experiment_name: 'rsitmd_LW_MCR' 57 | nums: 1 58 | current_num: 0 59 | -------------------------------------------------------------------------------- /predict/test_data/sparseresidential_3814.tif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xiaoyuan1996/SemanticLocalizationMetrics/1e9596ded8dbdb83ea534987f8c2e2c6ee3f1d20/predict/test_data/sparseresidential_3814.tif -------------------------------------------------------------------------------- /predict/util/convert_data.py: -------------------------------------------------------------------------------- 1 | # ----------------------------------------------------------- 2 | # Stacked Cross Attention Network implementation based on 3 | # https://arxiv.org/abs/1803.08024. 4 | # "Stacked Cross Attention for Image-Text Matching" 5 | # Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, Xiaodong He 6 | # 7 | # Writen by Kuang-Huei Lee, 2018 8 | # --------------------------------------------------------------- 9 | """Convert image features from bottom up attention to numpy array""" 10 | import argparse 11 | import base64 12 | import csv 13 | import os 14 | import sys 15 | 16 | import numpy as np 17 | 18 | parser = argparse.ArgumentParser() 19 | parser.add_argument('--imgid_list', default='../data/coco_precomp/train_ids.txt', 20 | help='Path to list of image id') 21 | parser.add_argument('--input_file', default=['../data/bu_data/trainval/karpathy_train_resnet101_faster_rcnn_genome.tsv.0'], 22 | # ,'../data/bu_data/trainval/karpathy_train_resnet101_faster_rcnn_genome.tsv.1'], 23 | help='tsv of all image data (output of bottom-up-attention/tools/generate_tsv.py), \ 24 | where each columns are: [image_id, image_w, image_h, num_boxes, boxes, features].') 25 | parser.add_argument('--output_dir', default='../data/coco_precomp/', 26 | help='Output directory.') 27 | parser.add_argument('--split', default='train', 28 | help='train|dev|test') 29 | opt = parser.parse_args() 30 | print(opt) 31 | 32 | 33 | meta = [] 34 | feature = {} 35 | for line in open(opt.imgid_list): 36 | sid = int(line.strip()) 37 | meta.append(sid) 38 | feature[sid] = None 39 | 40 | maxInt = sys.maxsize 41 | 42 | while True: 43 | # decrease the maxInt value by factor 10 44 | # as long as the OverflowError occurs. 45 | 46 | try: 47 | csv.field_size_limit(maxInt) 48 | break 49 | except OverflowError: 50 | maxInt = int(maxInt/10) 51 | 52 | FIELDNAMES = ['image_id', 'image_w', 'image_h', 'num_boxes', 'boxes', 'features'] 53 | 54 | if __name__ == '__main__': 55 | for input_file in opt.input_file: 56 | with open(input_file, "r+t") as tsv_in_file: 57 | reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES) 58 | for item in reader: 59 | item['image_id'] = int(item['image_id']) 60 | item['image_h'] = int(item['image_h']) 61 | item['image_w'] = int(item['image_w']) 62 | item['num_boxes'] = int(item['num_boxes']) 63 | for field in ['boxes', 'features']: 64 | item[field] = np.frombuffer(base64.decodestring(item[field].encode()), 65 | dtype=np.float32).reshape((item['num_boxes'],-1)) 66 | if item['image_id'] in feature: 67 | feature[item['image_id']] = item['features'] 68 | 69 | # Padding 70 | data_out = [] 71 | for sid in meta: 72 | padding_data = np.zeros((36,2048)) 73 | region_num = np.shape(feature[sid])[0] 74 | if region_num <= 36: 75 | padding_data[:region_num, :] = feature[sid] 76 | else: 77 | padding_data = feature[sid][:36, :] 78 | data_out.append(padding_data) 79 | 80 | data_out = np.stack(data_out, axis=0) 81 | print("Final numpy array shape:", data_out.shape) 82 | np.save(os.path.join(opt.output_dir, '{}_ims.npy'.format(opt.split)), data_out) -------------------------------------------------------------------------------- /predict/utils.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/05/05 6 | 7 | import logging 8 | import os 9 | import pickle 10 | import random 11 | 12 | import numpy as np 13 | import yaml 14 | 15 | 16 | # 加载参数 17 | def get_config(path_opt="common/config.yaml"): 18 | with open(path_opt, 'r') as handle: 19 | options = yaml.load(handle) 20 | return options 21 | 22 | # 储存为npy文件 23 | def save_to_npy(info, filename): 24 | np.save(filename,info,allow_pickle=True) 25 | 26 | # 从npy中读取 27 | def load_from_npy(filename): 28 | info = np.load(filename, allow_pickle=True) 29 | return info 30 | 31 | # 得到标准返回 32 | def get_stand_return(flag, message): 33 | code = 200 if flag else 400 34 | return_json = { 35 | 'code':code, 36 | "message": message 37 | } 38 | return return_json 39 | 40 | # 创建文件夹 41 | def create_dirs(path): 42 | if not os.path.exists(path): 43 | os.makedirs(path) 44 | 45 | # ========================== 字典操作 ============================= 46 | # 保存字典 47 | def dict_save(obj, name="rsd.pkl"): 48 | with open(name, 'wb') as f: 49 | pickle.dump(obj, f, pickle.HIGHEST_PROTOCOL) 50 | 51 | # 读取字典 52 | def dict_load(name="rsd.pkl"): 53 | with open(name, 'rb') as f: 54 | return pickle.load(f) 55 | 56 | # 删除键值 57 | def dict_delete(k, rsd): 58 | rsd.pop(k) 59 | return rsd 60 | 61 | # 插入键值 62 | def dict_insert(k, v, rsd): 63 | rsd[k] = v 64 | return rsd 65 | 66 | # 按照键值进行排序 输出排序好的键 67 | def sort_based_values(sims_dict): 68 | """ 69 | :param sims_dict: dict 70 | { 71 | '1': 0.2, 72 | "2": 0.4, 73 | "3": 0.3, 74 | "4": 0.23 75 | } 76 | :return: key 降序 ['2', '3', '4', '1'] 77 | """ 78 | sims_dict = sorted(sims_dict.items(), key=lambda item: item[1])[::-1] 79 | return [i[0] for i in sims_dict] 80 | 81 | 82 | # ============================ 加载或新建 retrieval system data ========================= 83 | def init_rsd(file_path): 84 | if os.path.exists(file_path): 85 | rsd = dict_load(file_path) 86 | else: 87 | rsd = {} 88 | return rsd 89 | 90 | # 删除文件夹及对应文件 91 | def delete_dire(dire): 92 | dir_list = [] 93 | for root, dirs, files in os.walk(dire): 94 | for afile in files: 95 | os.remove(os.path.join(root, afile)) 96 | for adir in dirs: 97 | dir_list.append(os.path.join(root, adir)) 98 | for bdir in dir_list: 99 | os.rmdir(bdir) 100 | 101 | # 保存结果到txt文件 102 | def log_to_txt( contexts=None,filename="save.txt", mark=False,encoding='UTF-8',add_n=False,mode='a'): 103 | f = open(filename, mode,encoding=encoding) 104 | if mark: 105 | sig = "------------------------------------------------\n" 106 | f.write(sig) 107 | elif isinstance(contexts, dict): 108 | tmp = "" 109 | for c in contexts.keys(): 110 | tmp += str(c)+" | "+ str(contexts[c]) +"\n" 111 | contexts = tmp 112 | f.write(contexts) 113 | else: 114 | if isinstance(contexts,list): 115 | tmp = "" 116 | for c in contexts: 117 | if add_n: 118 | tmp += str(c) + " " + "\n" 119 | else: 120 | tmp += str(c) + " " 121 | contexts = tmp 122 | else: 123 | contexts = contexts + "\n" 124 | f.write(contexts) 125 | 126 | 127 | f.close() 128 | 129 | # 从txt中读取行 130 | def load_from_txt(filename, encoding="utf-8"): 131 | f = open(filename,'r' ,encoding=encoding) 132 | contexts = f.readlines() 133 | return contexts 134 | 135 | # 创建随机文件夹 136 | def create_random_dirs_name(dir_path): 137 | dirs = os.listdir(dir_path) 138 | new_dir = "" 139 | while (new_dir == "") or (new_dir in dirs): 140 | new_dir = "".join(random.sample('1234567890qwertyuiopasdfghjklzxcvbnm', 8)) 141 | return new_dir 142 | 143 | # logger 144 | def get_logger(save_path=None): 145 | logger = logging.getLogger() 146 | logger.setLevel(logging.INFO) # 设置打印级别 147 | formatter = logging.Formatter('%(asctime)s %(message)s') 148 | 149 | # 设置屏幕打印的格式 150 | sh = logging.StreamHandler() 151 | sh.setFormatter(formatter) 152 | logger.addHandler(sh) 153 | 154 | # 设置log保存 155 | if save_path != None: 156 | fh = logging.FileHandler(save_path, encoding='utf8') 157 | fh.setFormatter(formatter) 158 | logger.addHandler(fh) 159 | 160 | return logger -------------------------------------------------------------------------------- /predict/vocab/rsitmd_splits_vocab.json: -------------------------------------------------------------------------------- 1 | {"word2idx": {"baseball": 1, "field": 2, "beside": 3, "green": 4, "amusement": 5, "park": 6, "around": 7, "red": 8, "track": 9, "adjacent": 10, "playground": 11, "square": 12, "long": 13, "path": 14, "next": 15, "runway": 16, "two": 17, "white": 18, "houses": 19, "located": 20, "middle": 21, "lawn": 22, "fields": 23, "different": 24, "sizes": 25, "stadium": 26, "stand": 27, "big": 28, "side-by-side": 29, "road": 30, "side": 31, "football": 32, "passes": 33, "gray": 34, "bare": 35, "soil": 36, "close": 37, "alongside": 38, "piece": 39, "land": 40, "separates": 41, "house": 42, "separating": 43, "octagonal": 44, "buildings": 45, "colors": 46, "building": 47, "parking": 48, "lot": 49, "full": 50, "cars": 51, "tennis": 52, "courts": 53, "grey": 54, "roadside": 55, "one": 56, "intersection": 57, "roads": 58, "crossed": 59, "highway": 60, "parallel": 61, "pass": 62, "streets": 63, "lined": 64, "trees": 65, "empty": 66, "sides": 67, "lots": 68, "deep": 69, "black": 70, "river": 71, "blue": 72, "factory": 73, "boats": 74, "moored": 75, "many": 76, "parked": 77, "bank": 78, "three": 79, "anchored": 80, "respectively": 81, "turquoise": 82, "ships": 83, "area": 84, "goods": 85, "boat": 86, "quietly": 87, "thin": 88, "bridge": 89, "sail": 90, "dark": 91, "sea": 92, "harbor": 93, "small": 94, "floating": 95, "port": 96, "floated": 97, "ship": 98, "near": 99, "dock": 100, "yellow": 101, "lay": 102, "across": 103, "crossroads": 104, "sailed": 105, "fast": 106, "calm": 107, "sailing": 108, "deck": 109, "purple": 110, "dry": 111, "cargo": 112, "striped": 113, "coast": 114, "stripes": 115, "shore": 116, "water": 117, "either": 118, "car": 119, "left": 120, "wave": 121, "heihe": 122, "pools": 123, "swimming": 124, "opposite": 125, "facing": 126, "moving": 127, "colorful": 128, "board": 129, "belt": 130, "curved": 131, "crosses": 132, "directly": 133, "make": 134, "way": 135, "residential": 136, "crossing": 137, "intersects": 138, "areas": 139, "plant": 140, "belts": 141, "rivers": 142, "surrounded": 143, "densely": 144, "populated": 145, "decorated": 146, "hand": 147, "open": 148, "grassland": 149, "high-rise": 150, "grass": 151, "skyscrapers": 152, "sits": 153, "tall": 154, "ellipse": 155, "shapes": 156, "oval": 157, "built": 158, "kind": 159, "light": 160, "bifurcation": 161, "street": 162, "train": 163, "railway": 164, "directions": 165, "tracks": 166, "planes": 167, "plane": 168, "large": 169, "crowded": 170, "separated": 171, "jagged": 172, "zigzag": 173, "lying": 174, "vehicles": 175, "airport": 176, "ground": 177, "stopped": 178, "dust": 179, "five": 180, "docked": 181, "sparkling": 182, "face": 183, "rectangular": 184, "together": 185, "line": 186, "several": 187, "edge": 188, "playgrounds": 189, "six": 190, "table": 191, "balls": 192, "semicircular": 193, "orange": 194, "dense": 195, "forest": 196, "thick": 197, "connected": 198, "lake": 199, "pond": 200, "lush": 201, "tree": 202, "connects": 203, "court": 204, "vacant": 205, "paths": 206, "expanse": 207, "run": 208, "runways": 209, "door": 210, "winding": 211, "covered": 212, "shaded": 213, "ponds": 214, "brown": 215, "spectators": 216, "spectator": 217, "audience": 218, "quite": 219, "lawns": 220, "divide": 221, "stood": 222, "four": 223, "bareland": 224, "leans": 225, "surround": 226, "front": 227, "terminal": 228, "others": 229, "besides": 230, "sparse": 231, "meadow": 232, "lines": 233, "inside": 234, "plants": 235, "neat": 236, "arranged": 237, "round": 238, "church": 239, "sit": 240, "aeroplanes": 241, "apron": 242, "flight": 243, "ploygon": 244, "main": 245, "netlike": 246, "cross": 247, "polygonal": 248, "passenger": 249, "termial": 250, "strip": 251, "shaped": 252, "viaducts": 253, "narrow": 254, "star": 255, "like": 256, "boomerang": 257, "aircraft": 258, "corner": 259, "take": 260, "running": 261, "seven": 262, "tarmac": 263, "boarding": 264, "gate": 265, "landside": 266, "airside": 267, "aircrafts": 268, "tarmacs": 269, "grassy": 270, "waiting": 271, "straight": 272, "standing": 273, "bar": 274, "airstrip": 275, "terminals": 276, "airplanes": 277, "circular": 278, "stretch": 279, "sparsely": 280, "rounded": 281, "rectangle": 282, "surrounds": 283, "radial": 284, "symmetrical": 285, "triangle": 286, "curve": 287, "seated": 288, "wasteland": 289, "polygon": 290, "irregular": 291, "paralleled": 292, "sandwiched": 293, "orderly": 294, "majestic": 295, "surrounding": 296, "arc": 297, "ways": 298, "pieces": 299, "meadows": 300, "strips": 301, "neatly": 302, "trapezoid": 303, "divided": 304, "outside": 305, "complicated": 306, "towers": 307, "cut": 308, "scattered": 309, "airplane": 310, "lamp": 311, "eight": 312, "sitting": 313, "pale": 314, "pure": 315, "y-shaped": 316, "berthing": 317, "conspicuous": 318, "beautiful": 319, "stands": 320, "especially": 321, "packed": 322, "silver": 323, "passed": 324, "wide": 325, "dozens": 326, "e-shaped": 327, "number": 328, "plain": 329, "seen": 330, "station": 331, "t-shaped": 332, "room": 333, "appeared": 334, "fragmented": 335, "terrain": 336, "nothing": 337, "stop": 338, "tower": 339, "blocks": 340, "composed": 341, "ball": 342, "u": 343, "smaller": 344, "single": 345, "farmland": 346, "simple": 347, "along": 348, "order": 349, "taxiing": 350, "rows": 351, "parts": 352, "corners": 353, "squared": 354, "stayed": 355, "nearby": 356, "traverse": 357, "contains": 358, "including": 359, "circle": 360, "ring": 361, "consists": 362, "arc-shaped": 363, "rectangles": 364, "separate": 365, "see": 366, "squares": 367, "wild": 368, "aprons": 369, "bright": 370, "block": 371, "facilities": 372, "complete": 373, "industrial": 374, "people": 375, "flat": 376, "spaces": 377, "huge": 378, "vast": 379, "without": 380, "reddish": 381, "meet": 382, "reddish-brown": 383, "obvious": 384, "marks": 385, "floor": 386, "tract": 387, "traces": 388, "clay": 389, "truck": 390, "driving": 391, "hill": 392, "mark": 393, "vegetation": 394, "khaki": 395, "'s": 396, "work": 397, "space": 398, "sand": 399, "leaking": 400, "looks": 401, "uneven": 402, "planted": 403, "surface": 404, "rugged": 405, "intersect": 406, "vertically": 407, "courtyard": 408, "patch": 409, "shape": 410, "interspersed": 411, "messy": 412, "undulating": 413, "wavy": 414, "sandy": 415, "dotted": 416, "bleak": 417, "dirt": 418, "desert": 419, "stretches": 420, "greenland": 421, "coner": 422, "form": 423, "rough": 424, "formed": 425, "fork": 426, "lie": 427, "traverses": 428, "shallow": 429, "clear": 430, "stripe": 431, "go": 432, "alleys": 433, "runs": 434, "earth": 435, "weeds": 436, "growing": 437, "waste": 438, "broken": 439, "stone": 440, "wheel": 441, "naked": 442, "country": 443, "place": 444, "grows": 445, "amount": 446, "roofs": 447, "kinds": 448, "thing": 449, "things": 450, "part": 451, "irregularly": 452, "blended": 453, "littered": 454, "dots": 455, "half": 456, "rock": 457, "hills": 458, "patterns": 459, "pattern": 460, "lands": 461, "use": 462, "wheels": 463, "holes": 464, "also": 465, "tree-lined": 466, "family": 467, "trucks": 468, "behind": 469, "barren": 470, "flowing": 471, "hillside": 472, "slope": 473, "center": 474, "forming": 475, "flower": 476, "luxuriant": 477, "flowers": 478, "row": 479, "top": 480, "view": 481, "gymnasium": 482, "circles": 483, "quarter": 484, "stadiums": 485, "fan-shaped": 486, "lakes": 487, "silvery": 488, "golf": 489, "course": 490, "filled": 491, "clean": 492, "places": 493, "enough": 494, "smooth": 495, "met": 496, "triangular": 497, "faces": 498, "perfect": 499, "rooms": 500, "players": 501, "lane": 502, "freeway": 503, "baseballfield": 504, "red-roofed": 505, "semi-surrounded": 506, "fan": 507, "fanshaped": 508, "reflections": 509, "distributed": 510, "occupies": 511, "pentagonal": 512, "basketball": 513, "forms": 514, "semicircle": 515, "baseballfields": 516, "another": 517, "sports": 518, "size": 519, "back": 520, "soccer": 521, "pitch": 522, "pitches": 523, "neighborhood": 524, "sized": 525, "highways": 526, "woods": 527, "vertical": 528, "routes": 529, "flanked": 530, "among": 531, "iron": 532, "solar": 533, "pool": 534, "grounds": 535, "bleachers": 536, "enclosed": 537, "coconut": 538, "edges": 539, "used": 540, "reflection": 541, "pentagon": 542, "devided": 543, "residence": 544, "lies": 545, "texture": 546, "items": 547, "addition": 548, "waves": 549, "ocean": 550, "beach": 551, "rocks": 552, "beat": 553, "saw": 554, "spectacular": 555, "magnificent": 556, "surging": 557, "spray": 558, "emerald": 559, "curves": 560, "oceans": 561, "float": 562, "patches": 563, "gradient": 564, "seems": 565, "color": 566, "prosperous": 567, "resort": 568, "thriving": 569, "fishing": 570, "ripples": 571, "look": 572, "forests": 573, "beaches": 574, "hit": 575, "picture": 576, "island": 577, "islands": 578, "quiet": 579, "high": 580, "blue-green": 581, "tourists": 582, "scenery": 583, "rocky": 584, "sands": 585, "right": 586, "dunes": 587, "soft": 588, "grasslands": 589, "grow": 590, "stones": 591, "series": 592, "coastline": 593, "dirty": 594, "waters": 595, "wet": 596, "rushed": 597, "city": 598, "layers": 599, "patting": 600, "wharf": 601, "living": 602, "mountain": 603, "short": 604, "boundary": 605, "thr": 606, "jungle": 607, "stained": 608, "dam": 609, "gently": 610, "ribbon": 611, "jade": 612, "ribbons": 613, "objects": 614, "meets": 615, "integrated": 616, "playing": 617, "play": 618, "sunshine": 619, "seaside": 620, "sun": 621, "connect": 622, "reaches": 623, "towards": 624, "golden": 625, "little": 626, "fish": 627, "found": 628, "yellowish": 629, "cyan": 630, "seas": 631, "umbrellas": 632, "yachts": 633, "better": 634, "stretching": 635, "extends": 636, "roofed": 637, "roof": 638, "shadows": 639, "shade": 640, "structure": 641, "scales": 642, "mountains": 643, "plots": 644, "groves": 645, "bottom": 646, "bridges": 647, "banks": 648, "flows": 649, "two-way": 650, "steel": 651, "made": 652, "spans": 653, "six-lane": 654, "frame": 655, "lanes": 656, "span": 657, "partially": 658, "types": 659, "broad": 660, "cast": 661, "shadow": 662, "muddy": 663, "similar": 664, "two-lane": 665, "turns": 666, "joins": 667, "crisscrossed": 668, "end": 669, "factories": 670, "bus": 671, "branches": 672, "traffic": 673, "yacht": 674, "decorations": 675, "pedestrians": 676, "various": 677, "spanning": 678, "connecting": 679, "valley": 680, "ones": 681, "plenty": 682, "bridege": 683, "suspension": 684, "tributary": 685, "cable": 686, "arch": 687, "divides": 688, "double": 689, "cities": 690, "farmlands": 691, "tiny": 692, "stopping": 693, "walking": 694, "bustling": 695, "low": 696, "saddle": 697, "confluence": 698, "turbid": 699, "goes": 700, "passing": 701, "traveling": 702, "cement": 703, "partly": 704, "riverbank": 705, "docks": 706, "non-parallel": 707, "slender": 708, "come": 709, "slowed": 710, "tanks": 711, "oil": 712, "intersections": 713, "villas": 714, "twists": 715, "reservoirs": 716, "storage": 717, "canal": 718, "wilderness": 719, "farm": 720, "arches": 721, "turning": 722, "villages": 723, "seperated": 724, "automobiles": 725, "tributaries": 726, "towns": 727, "diagonal": 728, "barelands": 729, "set": 730, "eyot": 731, "urban": 732, "ran": 733, "expressway": 734, "cruising": 735, "designed": 736, "design": 737, "conference": 738, "arcs": 739, "floors": 740, "convention": 741, "bend": 742, "tightly": 743, "centers": 744, "shell": 745, "intersecting": 746, "parks": 747, "wonderful": 748, "semi-circular": 749, "intensive": 750, "almost": 751, "crops": 752, "entrance": 753, "unique": 754, "central": 755, "away": 756, "memory": 757, "shows": 758, "concentric": 759, "public": 760, "plaza": 761, "trapezoidal": 762, "quadrilateral": 763, "leaf": 764, "eye": 765, "regular": 766, "hexagonal": 767, "rings": 768, "loop": 769, "hexagon": 770, "grid": 771, "lace": 772, "spherical": 773, "bowl": 774, "flyover": 775, "metal": 776, "esthetic": 777, "aesthetic": 778, "chess": 779, "silver-gray": 780, "nearly": 781, "incomplete": 782, "strange": 783, "grand": 784, "whole": 785, "vivid": 786, "gym": 787, "honeycomb": 788, "site": 789, "camp": 790, "pyramid": 791, "elliptical": 792, "construction": 793, "conical": 794, "glass": 795, "bushes": 796, "railways": 797, "placed": 798, "well": 799, "include": 800, "leaves": 801, "busy": 802, "layout": 803, "distribution": 804, "architectural": 805, "owns": 806, "centre": 807, "variety": 808, "region": 809, "styles": 810, "pink": 811, "moss": 812, "cathedral": 813, "great": 814, "dome": 815, "churches": 816, "spread": 817, "yard": 818, "properly": 819, "ordered": 820, "closed": 821, "semi": 822, "heart": 823, "mint": 824, "gree": 825, "pasture": 826, "every": 827, "includes": 828, "bricks": 829, "wooden": 830, "brick": 831, "garden": 832, "neighborhoods": 833, "junction": 834, "t-junction": 835, "old": 836, "flats": 837, "columns": 838, "cone": 839, "domes": 840, "sloping": 841, "cylindrical": 842, "structures": 843, "column": 844, "apartments": 845, "cruciform": 846, "redor": 847, "closely": 848, "linked": 849, "good": 850, "density": 851, "architecture": 852, "colored": 853, "bit": 854, "fine": 855, "umbrella": 856, "consisting": 857, "quadrangle": 858, "gathered": 859, "alley": 860, "wall": 861, "diamond": 862, "covers": 863, "blu": 864, "e": 865, "ceiling": 866, "flow": 867, "stream": 868, "mirror": 869, "reflected": 870, "bunch": 871, "average": 872, "rather": 873, "shrubs": 874, "nice": 875, "bent": 876, "commercial": 877, "business": 878, "district": 879, "concentrated": 880, "office": 881, "districts": 882, "downtown": 883, "well-organized": 884, "shopping": 885, "mixed": 886, "evergreen": 887, "numerous": 888, "home": 889, "section": 890, "viaduct": 891, "rare": 892, "travel": 893, "drive": 894, "skyscraper": 895, "ordinary": 896, "advertisements": 897, "compact": 898, "advertisement": 899, "rise": 900, "school": 901, "medium": 902, "apartment": 903, "edifices": 904, "height": 905, "overpass": 906, "relatively": 907, "standard": 908, "vary": 909, "distance": 910, "economic": 911, "developed": 912, "schools": 913, "mall": 914, "mainly": 915, "zone": 916, "split": 917, "major": 918, "commerce": 919, "beige": 920, "casting": 921, "billboard": 922, "community": 923, "sections": 924, "gardens": 925, "consist": 926, "sector": 927, "spot": 928, "convenient": 929, "territory": 930, "much": 931, "everywhere": 932, "bed": 933, "sharp": 934, "towering": 935, "villa": 936, "4th": 937, "encircle": 938, "serious": 939, "jams": 940, "nas": 941, "files": 942, "constitute": 943, "marked": 944, "range": 945, "organized": 946, "apart": 947, "compactly": 948, "wood": 949, "distribute": 950, "colours": 951, "pastures": 952, "aligned": 953, "cover": 954, "housing": 955, "shades": 956, "divids": 957, "crooked": 958, "oblique": 959, "came": 960, "rate": 961, "greening": 962, "growth": 963, "mostly": 964, "regularly": 965, "luxury": 966, "extending": 967, "never": 968, "stops": 969, "leisure": 970, "file": 971, "rooftop": 972, "village": 973, "trains": 974, "winds": 975, "homes": 976, "bem": 977, "tight": 978, "extend": 979, "arched": 980, "tem": 981, "agricultural": 982, "concrete": 983, "settled": 984, "tidy": 985, "angle": 986, "arrangement": 987, "past": 988, "regions": 989, "desolate": 990, "wrinkles": 991, "wrinkled": 992, "deserted": 993, "ca": 994, "n't": 995, "anything": 996, "oasis": 997, "deserts": 998, "cracks": 999, "scale": 1000, "dried": 1001, "figure": 1002, "stains": 1003, "layered": 1004, "pellets": 1005, "consistency": 1006, "visible": 1007, "grove": 1008, "curly": 1009, "crack": 1010, "spots": 1011, "roots": 1012, "tortuous": 1013, "thousands": 1014, "pile": 1015, "dusty": 1016, "arid": 1017, "boundless": 1018, "prairie": 1019, "endless": 1020, "resources": 1021, "nua": 1022, "scene": 1023, "dividing": 1024, "layer": 1025, "peaks": 1026, "band": 1027, "brook": 1028, "outline": 1029, "creamy": 1030, "cream": 1031, "pits": 1032, "snow": 1033, "complex": 1034, "textures": 1035, "petal": 1036, "colour": 1037, "rain": 1038, "really": 1039, "air": 1040, "fold": 1041, "wind": 1042, "gap": 1043, "looking": 1044, "steep": 1045, "turn": 1046, "dead": 1047, "trail": 1048, "alternating": 1049, "sidewalks": 1050, "farms": 1051, "coloured": 1052, "grown": 1053, "cultivated": 1054, "paddy": 1055, "rest": 1056, "ridge": 1057, "ridges": 1058, "landes": 1059, "mud": 1060, "depths": 1061, "crop": 1062, "loose": 1063, "checkered": 1064, "data": 1065, "points": 1066, "wheat": 1067, "leading": 1068, "luxurious": 1069, "uncultivated": 1070, "b": 1071, "forked": 1072, "bald": 1073, "polygons": 1074, "wooded": 1075, "town": 1076, "ramp": 1077, "nest": 1078, "clouds": 1079, "bush": 1080, "unusual": 1081, "families": 1082, "ranch": 1083, "landscape": 1084, "rich": 1085, "hut": 1086, "spring": 1087, "bottle": 1088, "footpath": 1089, "birds": 1090, "grandma": 1091, "flourishing": 1092, "ranges": 1093, "coverage": 1094, "clearly": 1095, "time": 1096, "twisted": 1097, "unpaved": 1098, "altitude": 1099, "tent": 1100, "zones": 1101, "tents": 1102, "v-shaped": 1103, "direction": 1104, "painted": 1105, "withered": 1106, "peak": 1107, "bud": 1108, "cloud": 1109, "shed": 1110, "laid": 1111, "industry": 1112, "workshops": 1113, "group": 1114, "workshop": 1115, "warehouse": 1116, "warehouses": 1117, "container": 1118, "containers": 1119, "paved": 1120, "betweeen": 1121, "barley": 1122, "cylinders": 1123, "basin": 1124, "modern": 1125, "nude": 1126, "belong": 1127, "tank": 1128, "rio": 1129, "whose": 1130, "overpasses": 1131, "valleys": 1132, "lively": 1133, "system": 1134, "pots": 1135, "cans": 1136, "gradually": 1137, "due": 1138, "problems": 1139, "turf": 1140, "prado": 1141, "perpendicular": 1142, "slightly": 1143, "worn": 1144, "boxes": 1145, "decorating": 1146, "bird": 1147, "folds": 1148, "sprouted": 1149, "yet": 1150, "loess": 1151, "fertile": 1152, "medium-sized": 1153, "hidden": 1154, "situated": 1155, "peaceful": 1156, "ten": 1157, "cottage": 1158, "cabins": 1159, "cottages": 1160, "smart": 1161, "residents": 1162, "walls": 1163, "u-shaped": 1164, "entertainment": 1165, "recreation": 1166, "independent": 1167, "still": 1168, "environment": 1169, "sidewalk": 1170, "venues": 1171, "forks": 1172, "leads": 1173, "rolling": 1174, "corrugated": 1175, "terraces": 1176, "eyes": 1177, "snows": 1178, "covering": 1179, "mountainous": 1180, "foot": 1181, "greek": 1182, "heart-shaped": 1183, "artificial": 1184, "lagoon": 1185, "ferris": 1186, "theme": 1187, "interesting": 1188, "outdoor": 1189, "interlaced": 1190, "equipped": 1191, "equipment": 1192, "children": 1193, "one-way": 1194, "seats": 1195, "position": 1196, "crammed": 1197, "hundreds": 1198, "asphalt": 1199, "spacious": 1200, "median": 1201, "ceilings": 1202, "pulled": 1203, "enter": 1204, "manner": 1205, "lost": 1206, "4": 1207, "vehicle": 1208, "platform": 1209, "goal": 1210, "crimson": 1211, "footballground": 1212, "game": 1213, "teaching": 1214, "educational": 1215, "jim": 1216, "twelve": 1217, "lead": 1218, "guys": 1219, "cemnet": 1220, "plastic": 1221, "intelligent": 1222, "hockey": 1223, "base": 1224, "south": 1225, "fica": 1226, "steps": 1227, "barefoot": 1228, "sky": 1229, "image": 1230, "fog": 1231, "mist": 1232, "diamonds": 1233, "combined": 1234, "spindle": 1235, "fountain": 1236, "stars": 1237, "takes": 1238, "l-shaped": 1239, "2": 1240, "ellipses": 1241, "ports": 1242, "harbour": 1243, "pier": 1244, "quality": 1245, "reach": 1246, "docking": 1247, "seawall": 1248, "basins": 1249, "sailboats": 1250, "regulated": 1251, "leave": 1252, "stations": 1253, "rail": 1254, "railroad": 1255, "facility": 1256, "spindle-shaped": 1257, "satellite": 1258, "awnings": 1259, "awning": 1260, "high-speed": 1261, "belongs": 1262, "subway": 1263, "gray-roofed": 1264, "garage": 1265, "guide": 1266, "transportation": 1267, "resorts": 1268, "tropical": 1269, "tourist": 1270, "ahead": 1271, "holiday": 1272, "chairs": 1273, "s-shaped": 1274, "palm": 1275, "roundabout": 1276, "herringbone": 1277, "rainforest": 1278, "wetlands": 1279, "flowed": 1280, "join": 1281, "saddle-shaped": 1282, "greenbelts": 1283, "cake": 1284, "campus": 1285, "northeast": 1286, "countless": 1287, "painting": 1288, "university": 1289, "northwest": 1290, "well-equipped": 1291, "students": 1292, "carefully": 1293, "c-shaped": 1294, "auditorium": 1295, "hot": 1296, "private": 1297, "photo": 1298, "beds": 1299, "hall": 1300, "monument": 1301, "lantern": 1302, "bell": 1303, "x-shaped": 1304, "hold": 1305, "tens": 1306, "new": 1307, "tripods": 1308, "stalls": 1309, "blearchers": 1310, "pipe": 1311, "storagetanks": 1312, "pipelines": 1313, "drums": 1314, "pipes": 1315, "reservoir": 1316, "identical": 1317, "pallets": 1318, "nine": 1319, "trays": 1320, "storerooms": 1321, "barrelan": 1322, "june": 1323, "columnar": 1324, "tanker": 1325, "loaded": 1326, "treatment": 1327, "jars": 1328, "refinery": 1329, "tankers": 1330, "storing": 1331, "jar": 1332, "bicycles": 1333, "barrels": 1334, "shrinking": 1335, "loops": 1336, "auxiliary": 1337, "eight-shaped": 1338, "flyovers": 1339, "ramps": 1340, "three-dimensional": 1341, "brush": 1342, "nested": 1343, "luggage": 1344, "turfs": 1345, "bunkers": 1346, "extremely": 1347, "20": 1348, "": 1349}, "idx2word": {"1": "baseball", "2": "field", "3": "beside", "4": "green", "5": "amusement", "6": "park", "7": "around", "8": "red", "9": "track", "10": "adjacent", "11": "playground", "12": "square", "13": "long", "14": "path", "15": "next", "16": "runway", "17": "two", "18": "white", "19": "houses", "20": "located", "21": "middle", "22": "lawn", "23": "fields", "24": "different", "25": "sizes", "26": "stadium", "27": "stand", "28": "big", "29": "side-by-side", "30": "road", "31": "side", "32": "football", "33": "passes", "34": "gray", "35": "bare", "36": "soil", "37": "close", "38": "alongside", "39": "piece", "40": "land", "41": "separates", "42": "house", "43": "separating", "44": "octagonal", "45": "buildings", "46": "colors", "47": "building", "48": "parking", "49": "lot", "50": "full", "51": "cars", "52": "tennis", "53": "courts", "54": "grey", "55": "roadside", "56": "one", "57": "intersection", "58": "roads", "59": "crossed", "60": "highway", "61": "parallel", "62": "pass", "63": "streets", "64": "lined", "65": "trees", "66": "empty", "67": "sides", "68": "lots", "69": "deep", "70": "black", "71": "river", "72": "blue", "73": "factory", "74": "boats", "75": "moored", "76": "many", "77": "parked", "78": "bank", "79": "three", "80": "anchored", "81": "respectively", "82": "turquoise", "83": "ships", "84": "area", "85": "goods", "86": "boat", "87": "quietly", "88": "thin", "89": "bridge", "90": "sail", "91": "dark", "92": "sea", "93": "harbor", "94": "small", "95": "floating", "96": "port", "97": "floated", "98": "ship", "99": "near", "100": "dock", "101": "yellow", "102": "lay", "103": "across", "104": "crossroads", "105": "sailed", "106": "fast", "107": "calm", "108": "sailing", "109": "deck", "110": "purple", "111": "dry", "112": "cargo", "113": "striped", "114": "coast", "115": "stripes", "116": "shore", "117": "water", "118": "either", "119": "car", "120": "left", "121": "wave", "122": "heihe", "123": "pools", "124": "swimming", "125": "opposite", "126": "facing", "127": "moving", "128": "colorful", "129": "board", "130": "belt", "131": "curved", "132": "crosses", "133": "directly", "134": "make", "135": "way", "136": "residential", "137": "crossing", "138": "intersects", "139": "areas", "140": "plant", "141": "belts", "142": "rivers", "143": "surrounded", "144": "densely", "145": "populated", "146": "decorated", "147": "hand", "148": "open", "149": "grassland", "150": "high-rise", "151": "grass", "152": "skyscrapers", "153": "sits", "154": "tall", "155": "ellipse", "156": "shapes", "157": "oval", "158": "built", "159": "kind", "160": "light", "161": "bifurcation", "162": "street", "163": "train", "164": "railway", "165": "directions", "166": "tracks", "167": "planes", "168": "plane", "169": "large", "170": "crowded", "171": "separated", "172": "jagged", "173": "zigzag", "174": "lying", "175": "vehicles", "176": "airport", "177": "ground", "178": "stopped", "179": "dust", "180": "five", "181": "docked", "182": "sparkling", "183": "face", "184": "rectangular", "185": "together", "186": "line", "187": "several", "188": "edge", "189": "playgrounds", "190": "six", "191": "table", "192": "balls", "193": "semicircular", "194": "orange", "195": "dense", "196": "forest", "197": "thick", "198": "connected", "199": "lake", "200": "pond", "201": "lush", "202": "tree", "203": "connects", "204": "court", "205": "vacant", "206": "paths", "207": "expanse", "208": "run", "209": "runways", "210": "door", "211": "winding", "212": "covered", "213": "shaded", "214": "ponds", "215": "brown", "216": "spectators", "217": "spectator", "218": "audience", "219": "quite", "220": "lawns", "221": "divide", "222": "stood", "223": "four", "224": "bareland", "225": "leans", "226": "surround", "227": "front", "228": "terminal", "229": "others", "230": "besides", "231": "sparse", "232": "meadow", "233": "lines", "234": "inside", "235": "plants", "236": "neat", "237": "arranged", "238": "round", "239": "church", "240": "sit", "241": "aeroplanes", "242": "apron", "243": "flight", "244": "ploygon", "245": "main", "246": "netlike", "247": "cross", "248": "polygonal", "249": "passenger", "250": "termial", "251": "strip", "252": "shaped", "253": "viaducts", "254": "narrow", "255": "star", "256": "like", "257": "boomerang", "258": "aircraft", "259": "corner", "260": "take", "261": "running", "262": "seven", "263": "tarmac", "264": "boarding", "265": "gate", "266": "landside", "267": "airside", "268": "aircrafts", "269": "tarmacs", "270": "grassy", "271": "waiting", "272": "straight", "273": "standing", "274": "bar", "275": "airstrip", "276": "terminals", "277": "airplanes", "278": "circular", "279": "stretch", "280": "sparsely", "281": "rounded", "282": "rectangle", "283": "surrounds", "284": "radial", "285": "symmetrical", "286": "triangle", "287": "curve", "288": "seated", "289": "wasteland", "290": "polygon", "291": "irregular", "292": "paralleled", "293": "sandwiched", "294": "orderly", "295": "majestic", "296": "surrounding", "297": "arc", "298": "ways", "299": "pieces", "300": "meadows", "301": "strips", "302": "neatly", "303": "trapezoid", "304": "divided", "305": "outside", "306": "complicated", "307": "towers", "308": "cut", "309": "scattered", "310": "airplane", "311": "lamp", "312": "eight", "313": "sitting", "314": "pale", "315": "pure", "316": "y-shaped", "317": "berthing", "318": "conspicuous", "319": "beautiful", "320": "stands", "321": "especially", "322": "packed", "323": "silver", "324": "passed", "325": "wide", "326": "dozens", "327": "e-shaped", "328": "number", "329": "plain", "330": "seen", "331": "station", "332": "t-shaped", "333": "room", "334": "appeared", "335": "fragmented", "336": "terrain", "337": "nothing", "338": "stop", "339": "tower", "340": "blocks", "341": "composed", "342": "ball", "343": "u", "344": "smaller", "345": "single", "346": "farmland", "347": "simple", "348": "along", "349": "order", "350": "taxiing", "351": "rows", "352": "parts", "353": "corners", "354": "squared", "355": "stayed", "356": "nearby", "357": "traverse", "358": "contains", "359": "including", "360": "circle", "361": "ring", "362": "consists", "363": "arc-shaped", "364": "rectangles", "365": "separate", "366": "see", "367": "squares", "368": "wild", "369": "aprons", "370": "bright", "371": "block", "372": "facilities", "373": "complete", "374": "industrial", "375": "people", "376": "flat", "377": "spaces", "378": "huge", "379": "vast", "380": "without", "381": "reddish", "382": "meet", "383": "reddish-brown", "384": "obvious", "385": "marks", "386": "floor", "387": "tract", "388": "traces", "389": "clay", "390": "truck", "391": "driving", "392": "hill", "393": "mark", "394": "vegetation", "395": "khaki", "396": "'s", "397": "work", "398": "space", "399": "sand", "400": "leaking", "401": "looks", "402": "uneven", "403": "planted", "404": "surface", "405": "rugged", "406": "intersect", "407": "vertically", "408": "courtyard", "409": "patch", "410": "shape", "411": "interspersed", "412": "messy", "413": "undulating", "414": "wavy", "415": "sandy", "416": "dotted", "417": "bleak", "418": "dirt", "419": "desert", "420": "stretches", "421": "greenland", "422": "coner", "423": "form", "424": "rough", "425": "formed", "426": "fork", "427": "lie", "428": "traverses", "429": "shallow", "430": "clear", "431": "stripe", "432": "go", "433": "alleys", "434": "runs", "435": "earth", "436": "weeds", "437": "growing", "438": "waste", "439": "broken", "440": "stone", "441": "wheel", "442": "naked", "443": "country", "444": "place", "445": "grows", "446": "amount", "447": "roofs", "448": "kinds", "449": "thing", "450": "things", "451": "part", "452": "irregularly", "453": "blended", "454": "littered", "455": "dots", "456": "half", "457": "rock", "458": "hills", "459": "patterns", "460": "pattern", "461": "lands", "462": "use", "463": "wheels", "464": "holes", "465": "also", "466": "tree-lined", "467": "family", "468": "trucks", "469": "behind", "470": "barren", "471": "flowing", "472": "hillside", "473": "slope", "474": "center", "475": "forming", "476": "flower", "477": "luxuriant", "478": "flowers", "479": "row", "480": "top", "481": "view", "482": "gymnasium", "483": "circles", "484": "quarter", "485": "stadiums", "486": "fan-shaped", "487": "lakes", "488": "silvery", "489": "golf", "490": "course", "491": "filled", "492": "clean", "493": "places", "494": "enough", "495": "smooth", "496": "met", "497": "triangular", "498": "faces", "499": "perfect", "500": "rooms", "501": "players", "502": "lane", "503": "freeway", "504": "baseballfield", "505": "red-roofed", "506": "semi-surrounded", "507": "fan", "508": "fanshaped", "509": "reflections", "510": "distributed", "511": "occupies", "512": "pentagonal", "513": "basketball", "514": "forms", "515": "semicircle", "516": "baseballfields", "517": "another", "518": "sports", "519": "size", "520": "back", "521": "soccer", "522": "pitch", "523": "pitches", "524": "neighborhood", "525": "sized", "526": "highways", "527": "woods", "528": "vertical", "529": "routes", "530": "flanked", "531": "among", "532": "iron", "533": "solar", "534": "pool", "535": "grounds", "536": "bleachers", "537": "enclosed", "538": "coconut", "539": "edges", "540": "used", "541": "reflection", "542": "pentagon", "543": "devided", "544": "residence", "545": "lies", "546": "texture", "547": "items", "548": "addition", "549": "waves", "550": "ocean", "551": "beach", "552": "rocks", "553": "beat", "554": "saw", "555": "spectacular", "556": "magnificent", "557": "surging", "558": "spray", "559": "emerald", "560": "curves", "561": "oceans", "562": "float", "563": "patches", "564": "gradient", "565": "seems", "566": "color", "567": "prosperous", "568": "resort", "569": "thriving", "570": "fishing", "571": "ripples", "572": "look", "573": "forests", "574": "beaches", "575": "hit", "576": "picture", "577": "island", "578": "islands", "579": "quiet", "580": "high", "581": "blue-green", "582": "tourists", "583": "scenery", "584": "rocky", "585": "sands", "586": "right", "587": "dunes", "588": "soft", "589": "grasslands", "590": "grow", "591": "stones", "592": "series", "593": "coastline", "594": "dirty", "595": "waters", "596": "wet", "597": "rushed", "598": "city", "599": "layers", "600": "patting", "601": "wharf", "602": "living", "603": "mountain", "604": "short", "605": "boundary", "606": "thr", "607": "jungle", "608": "stained", "609": "dam", "610": "gently", "611": "ribbon", "612": "jade", "613": "ribbons", "614": "objects", "615": "meets", "616": "integrated", "617": "playing", "618": "play", "619": "sunshine", "620": "seaside", "621": "sun", "622": "connect", "623": "reaches", "624": "towards", "625": "golden", "626": "little", "627": "fish", "628": "found", "629": "yellowish", "630": "cyan", "631": "seas", "632": "umbrellas", "633": "yachts", "634": "better", "635": "stretching", "636": "extends", "637": "roofed", "638": "roof", "639": "shadows", "640": "shade", "641": "structure", "642": "scales", "643": "mountains", "644": "plots", "645": "groves", "646": "bottom", "647": "bridges", "648": "banks", "649": "flows", "650": "two-way", "651": "steel", "652": "made", "653": "spans", "654": "six-lane", "655": "frame", "656": "lanes", "657": "span", "658": "partially", "659": "types", "660": "broad", "661": "cast", "662": "shadow", "663": "muddy", "664": "similar", "665": "two-lane", "666": "turns", "667": "joins", "668": "crisscrossed", "669": "end", "670": "factories", "671": "bus", "672": "branches", "673": "traffic", "674": "yacht", "675": "decorations", "676": "pedestrians", "677": "various", "678": "spanning", "679": "connecting", "680": "valley", "681": "ones", "682": "plenty", "683": "bridege", "684": "suspension", "685": "tributary", "686": "cable", "687": "arch", "688": "divides", "689": "double", "690": "cities", "691": "farmlands", "692": "tiny", "693": "stopping", "694": "walking", "695": "bustling", "696": "low", "697": "saddle", "698": "confluence", "699": "turbid", "700": "goes", "701": "passing", "702": "traveling", "703": "cement", "704": "partly", "705": "riverbank", "706": "docks", "707": "non-parallel", "708": "slender", "709": "come", "710": "slowed", "711": "tanks", "712": "oil", "713": "intersections", "714": "villas", "715": "twists", "716": "reservoirs", "717": "storage", "718": "canal", "719": "wilderness", "720": "farm", "721": "arches", "722": "turning", "723": "villages", "724": "seperated", "725": "automobiles", "726": "tributaries", "727": "towns", "728": "diagonal", "729": "barelands", "730": "set", "731": "eyot", "732": "urban", "733": "ran", "734": "expressway", "735": "cruising", "736": "designed", "737": "design", "738": "conference", "739": "arcs", "740": "floors", "741": "convention", "742": "bend", "743": "tightly", "744": "centers", "745": "shell", "746": "intersecting", "747": "parks", "748": "wonderful", "749": "semi-circular", "750": "intensive", "751": "almost", "752": "crops", "753": "entrance", "754": "unique", "755": "central", "756": "away", "757": "memory", "758": "shows", "759": "concentric", "760": "public", "761": "plaza", "762": "trapezoidal", "763": "quadrilateral", "764": "leaf", "765": "eye", "766": "regular", "767": "hexagonal", "768": "rings", "769": "loop", "770": "hexagon", "771": "grid", "772": "lace", "773": "spherical", "774": "bowl", "775": "flyover", "776": "metal", "777": "esthetic", "778": "aesthetic", "779": "chess", "780": "silver-gray", "781": "nearly", "782": "incomplete", "783": "strange", "784": "grand", "785": "whole", "786": "vivid", "787": "gym", "788": "honeycomb", "789": "site", "790": "camp", "791": "pyramid", "792": "elliptical", "793": "construction", "794": "conical", "795": "glass", "796": "bushes", "797": "railways", "798": "placed", "799": "well", "800": "include", "801": "leaves", "802": "busy", "803": "layout", "804": "distribution", "805": "architectural", "806": "owns", "807": "centre", "808": "variety", "809": "region", "810": "styles", "811": "pink", "812": "moss", "813": "cathedral", "814": "great", "815": "dome", "816": "churches", "817": "spread", "818": "yard", "819": "properly", "820": "ordered", "821": "closed", "822": "semi", "823": "heart", "824": "mint", "825": "gree", "826": "pasture", "827": "every", "828": "includes", "829": "bricks", "830": "wooden", "831": "brick", "832": "garden", "833": "neighborhoods", "834": "junction", "835": "t-junction", "836": "old", "837": "flats", "838": "columns", "839": "cone", "840": "domes", "841": "sloping", "842": "cylindrical", "843": "structures", "844": "column", "845": "apartments", "846": "cruciform", "847": "redor", "848": "closely", "849": "linked", "850": "good", "851": "density", "852": "architecture", "853": "colored", "854": "bit", "855": "fine", "856": "umbrella", "857": "consisting", "858": "quadrangle", "859": "gathered", "860": "alley", "861": "wall", "862": "diamond", "863": "covers", "864": "blu", "865": "e", "866": "ceiling", "867": "flow", "868": "stream", "869": "mirror", "870": "reflected", "871": "bunch", "872": "average", "873": "rather", "874": "shrubs", "875": "nice", "876": "bent", "877": "commercial", "878": "business", "879": "district", "880": "concentrated", "881": "office", "882": "districts", "883": "downtown", "884": "well-organized", "885": "shopping", "886": "mixed", "887": "evergreen", "888": "numerous", "889": "home", "890": "section", "891": "viaduct", "892": "rare", "893": "travel", "894": "drive", "895": "skyscraper", "896": "ordinary", "897": "advertisements", "898": "compact", "899": "advertisement", "900": "rise", "901": "school", "902": "medium", "903": "apartment", "904": "edifices", "905": "height", "906": "overpass", "907": "relatively", "908": "standard", "909": "vary", "910": "distance", "911": "economic", "912": "developed", "913": "schools", "914": "mall", "915": "mainly", "916": "zone", "917": "split", "918": "major", "919": "commerce", "920": "beige", "921": "casting", "922": "billboard", "923": "community", "924": "sections", "925": "gardens", "926": "consist", "927": "sector", "928": "spot", "929": "convenient", "930": "territory", "931": "much", "932": "everywhere", "933": "bed", "934": "sharp", "935": "towering", "936": "villa", "937": "4th", "938": "encircle", "939": "serious", "940": "jams", "941": "nas", "942": "files", "943": "constitute", "944": "marked", "945": "range", "946": "organized", "947": "apart", "948": "compactly", "949": "wood", "950": "distribute", "951": "colours", "952": "pastures", "953": "aligned", "954": "cover", "955": "housing", "956": "shades", "957": "divids", "958": "crooked", "959": "oblique", "960": "came", "961": "rate", "962": "greening", "963": "growth", "964": "mostly", "965": "regularly", "966": "luxury", "967": "extending", "968": "never", "969": "stops", "970": "leisure", "971": "file", "972": "rooftop", "973": "village", "974": "trains", "975": "winds", "976": "homes", "977": "bem", "978": "tight", "979": "extend", "980": "arched", "981": "tem", "982": "agricultural", "983": "concrete", "984": "settled", "985": "tidy", "986": "angle", "987": "arrangement", "988": "past", "989": "regions", "990": "desolate", "991": "wrinkles", "992": "wrinkled", "993": "deserted", "994": "ca", "995": "n't", "996": "anything", "997": "oasis", "998": "deserts", "999": "cracks", "1000": "scale", "1001": "dried", "1002": "figure", "1003": "stains", "1004": "layered", "1005": "pellets", "1006": "consistency", "1007": "visible", "1008": "grove", "1009": "curly", "1010": "crack", "1011": "spots", "1012": "roots", "1013": "tortuous", "1014": "thousands", "1015": "pile", "1016": "dusty", "1017": "arid", "1018": "boundless", "1019": "prairie", "1020": "endless", "1021": "resources", "1022": "nua", "1023": "scene", "1024": "dividing", "1025": "layer", "1026": "peaks", "1027": "band", "1028": "brook", "1029": "outline", "1030": "creamy", "1031": "cream", "1032": "pits", "1033": "snow", "1034": "complex", "1035": "textures", "1036": "petal", "1037": "colour", "1038": "rain", "1039": "really", "1040": "air", "1041": "fold", "1042": "wind", "1043": "gap", "1044": "looking", "1045": "steep", "1046": "turn", "1047": "dead", "1048": "trail", "1049": "alternating", "1050": "sidewalks", "1051": "farms", "1052": "coloured", "1053": "grown", "1054": "cultivated", "1055": "paddy", "1056": "rest", "1057": "ridge", "1058": "ridges", "1059": "landes", "1060": "mud", "1061": "depths", "1062": "crop", "1063": "loose", "1064": "checkered", "1065": "data", "1066": "points", "1067": "wheat", "1068": "leading", "1069": "luxurious", "1070": "uncultivated", "1071": "b", "1072": "forked", "1073": "bald", "1074": "polygons", "1075": "wooded", "1076": "town", "1077": "ramp", "1078": "nest", "1079": "clouds", "1080": "bush", "1081": "unusual", "1082": "families", "1083": "ranch", "1084": "landscape", "1085": "rich", "1086": "hut", "1087": "spring", "1088": "bottle", "1089": "footpath", "1090": "birds", "1091": "grandma", "1092": "flourishing", "1093": "ranges", "1094": "coverage", "1095": "clearly", "1096": "time", "1097": "twisted", "1098": "unpaved", "1099": "altitude", "1100": "tent", "1101": "zones", "1102": "tents", "1103": "v-shaped", "1104": "direction", "1105": "painted", "1106": "withered", "1107": "peak", "1108": "bud", "1109": "cloud", "1110": "shed", "1111": "laid", "1112": "industry", "1113": "workshops", "1114": "group", "1115": "workshop", "1116": "warehouse", "1117": "warehouses", "1118": "container", "1119": "containers", "1120": "paved", "1121": "betweeen", "1122": "barley", "1123": "cylinders", "1124": "basin", "1125": "modern", "1126": "nude", "1127": "belong", "1128": "tank", "1129": "rio", "1130": "whose", "1131": "overpasses", "1132": "valleys", "1133": "lively", "1134": "system", "1135": "pots", "1136": "cans", "1137": "gradually", "1138": "due", "1139": "problems", "1140": "turf", "1141": "prado", "1142": "perpendicular", "1143": "slightly", "1144": "worn", "1145": "boxes", "1146": "decorating", "1147": "bird", "1148": "folds", "1149": "sprouted", "1150": "yet", "1151": "loess", "1152": "fertile", "1153": "medium-sized", "1154": "hidden", "1155": "situated", "1156": "peaceful", "1157": "ten", "1158": "cottage", "1159": "cabins", "1160": "cottages", "1161": "smart", "1162": "residents", "1163": "walls", "1164": "u-shaped", "1165": "entertainment", "1166": "recreation", "1167": "independent", "1168": "still", "1169": "environment", "1170": "sidewalk", "1171": "venues", "1172": "forks", "1173": "leads", "1174": "rolling", "1175": "corrugated", "1176": "terraces", "1177": "eyes", "1178": "snows", "1179": "covering", "1180": "mountainous", "1181": "foot", "1182": "greek", "1183": "heart-shaped", "1184": "artificial", "1185": "lagoon", "1186": "ferris", "1187": "theme", "1188": "interesting", "1189": "outdoor", "1190": "interlaced", "1191": "equipped", "1192": "equipment", "1193": "children", "1194": "one-way", "1195": "seats", "1196": "position", "1197": "crammed", "1198": "hundreds", "1199": "asphalt", "1200": "spacious", "1201": "median", "1202": "ceilings", "1203": "pulled", "1204": "enter", "1205": "manner", "1206": "lost", "1207": "4", "1208": "vehicle", "1209": "platform", "1210": "goal", "1211": "crimson", "1212": "footballground", "1213": "game", "1214": "teaching", "1215": "educational", "1216": "jim", "1217": "twelve", "1218": "lead", "1219": "guys", "1220": "cemnet", "1221": "plastic", "1222": "intelligent", "1223": "hockey", "1224": "base", "1225": "south", "1226": "fica", "1227": "steps", "1228": "barefoot", "1229": "sky", "1230": "image", "1231": "fog", "1232": "mist", "1233": "diamonds", "1234": "combined", "1235": "spindle", "1236": "fountain", "1237": "stars", "1238": "takes", "1239": "l-shaped", "1240": "2", "1241": "ellipses", "1242": "ports", "1243": "harbour", "1244": "pier", "1245": "quality", "1246": "reach", "1247": "docking", "1248": "seawall", "1249": "basins", "1250": "sailboats", "1251": "regulated", "1252": "leave", "1253": "stations", "1254": "rail", "1255": "railroad", "1256": "facility", "1257": "spindle-shaped", "1258": "satellite", "1259": "awnings", "1260": "awning", "1261": "high-speed", "1262": "belongs", "1263": "subway", "1264": "gray-roofed", "1265": "garage", "1266": "guide", "1267": "transportation", "1268": "resorts", "1269": "tropical", "1270": "tourist", "1271": "ahead", "1272": "holiday", "1273": "chairs", "1274": "s-shaped", "1275": "palm", "1276": "roundabout", "1277": "herringbone", "1278": "rainforest", "1279": "wetlands", "1280": "flowed", "1281": "join", "1282": "saddle-shaped", "1283": "greenbelts", "1284": "cake", "1285": "campus", "1286": "northeast", "1287": "countless", "1288": "painting", "1289": "university", "1290": "northwest", "1291": "well-equipped", "1292": "students", "1293": "carefully", "1294": "c-shaped", "1295": "auditorium", "1296": "hot", "1297": "private", "1298": "photo", "1299": "beds", "1300": "hall", "1301": "monument", "1302": "lantern", "1303": "bell", "1304": "x-shaped", "1305": "hold", "1306": "tens", "1307": "new", "1308": "tripods", "1309": "stalls", "1310": "blearchers", "1311": "pipe", "1312": "storagetanks", "1313": "pipelines", "1314": "drums", "1315": "pipes", "1316": "reservoir", "1317": "identical", "1318": "pallets", "1319": "nine", "1320": "trays", "1321": "storerooms", "1322": "barrelan", "1323": "june", "1324": "columnar", "1325": "tanker", "1326": "loaded", "1327": "treatment", "1328": "jars", "1329": "refinery", "1330": "tankers", "1331": "storing", "1332": "jar", "1333": "bicycles", "1334": "barrels", "1335": "shrinking", "1336": "loops", "1337": "auxiliary", "1338": "eight-shaped", "1339": "flyovers", "1340": "ramps", "1341": "three-dimensional", "1342": "brush", "1343": "nested", "1344": "luggage", "1345": "turfs", "1346": "bunkers", "1347": "extremely", "1348": "20", "1349": ""}, "idx": 1350} -------------------------------------------------------------------------------- /predict/vocabs.py: -------------------------------------------------------------------------------- 1 | # ----------------------------------------------------------- 2 | # "Remote Sensing Cross-Modal Text-Image Retrieval Based on Global and Local Information" 3 | # Yuan, Zhiqiang and Zhang, Wenkai and Changyuan Tian and Xuee, Rong and Zhengyuan Zhang and Wang, Hongqi and Fu, Kun and Sun, Xian 4 | # Writen by YuanZhiqiang, 2021. Our code is depended on AMFMN 5 | # ------------------------------------------------------------ 6 | 7 | import argparse 8 | import json 9 | import os 10 | from collections import Counter 11 | 12 | import nltk 13 | 14 | annotations = { 15 | 'coco_splits': ['train_caps.txt', 'val_caps.txt', 'test_caps.txt'], 16 | 'flickr30k_splits': ['train_caps.txt', 'val_caps.txt', 'test_caps.txt'], 17 | 'rsicd_precomp': ['train_caps.txt', 'val_caps.txt', 'test_caps.txt'], 18 | 'rsitmd_precomp': ['train_caps.txt', 'val_caps.txt'], 19 | 'ucm_precomp': ['train_caps.txt', 'val_caps.txt'], 20 | 'sydney_precomp': ['train_caps.txt', 'val_caps.txt'], 21 | 22 | } 23 | 24 | 25 | class Vocabulary(object): 26 | """Simple vocabulary wrapper.""" 27 | 28 | def __init__(self): 29 | self.word2idx = {} 30 | self.idx2word = {} 31 | self.idx = 1 32 | 33 | def add_word(self, word): 34 | if word not in self.word2idx: 35 | self.word2idx[word] = self.idx 36 | self.idx2word[self.idx] = word 37 | self.idx += 1 38 | 39 | def __call__(self, word): 40 | if word not in self.word2idx: 41 | return self.word2idx[''] 42 | return self.word2idx[word] 43 | 44 | def __len__(self): 45 | return len(self.word2idx) 46 | 47 | 48 | def serialize_vocab(vocab, dest): 49 | d = {} 50 | d['word2idx'] = vocab.word2idx 51 | d['idx2word'] = vocab.idx2word 52 | d['idx'] = vocab.idx 53 | with open(dest, "w") as f: 54 | json.dump(d, f) 55 | 56 | 57 | def deserialize_vocab(src): 58 | with open(src) as f: 59 | d = json.load(f) 60 | vocab = Vocabulary() 61 | vocab.word2idx = d['word2idx'] 62 | vocab.idx2word = d['idx2word'] 63 | vocab.idx = d['idx'] 64 | return vocab 65 | 66 | 67 | def from_txt(txt): 68 | captions = [] 69 | with open(txt, 'rb') as f: 70 | for line in f: 71 | captions.append(line.strip()) 72 | return captions 73 | 74 | 75 | def build_vocab(data_path, data_name, caption_file, threshold): 76 | """Build a simple vocabulary wrapper.""" 77 | 78 | stopword_list = list(set(nltk.corpus.stopwords.words('english'))) 79 | counter = Counter() 80 | for path in caption_file[data_name]: 81 | full_path = os.path.join(os.path.join(data_path, data_name), path) 82 | captions = from_txt(full_path) 83 | 84 | for i, caption in enumerate(captions): 85 | tokens = nltk.tokenize.word_tokenize( 86 | caption.lower().decode('utf-8')) 87 | punctuations = [',', '.', ':', ';', '?', '(', ')', '[', ']', '&', '!', '*', '@', '#', '$', '%'] 88 | tokens = [k for k in tokens if k not in punctuations] 89 | tokens = [k for k in tokens if k not in stopword_list] 90 | counter.update(tokens) 91 | 92 | if i % 1000 == 0: 93 | print("[%d/%d] tokenized the captions." % (i, len(captions))) 94 | 95 | # Discard if the occurrence of the word is less than min_word_cnt. 96 | words = [word for word, cnt in counter.items() if cnt >= threshold] 97 | 98 | # Create a vocab wrapper and add some special tokens. 99 | vocab = Vocabulary() 100 | 101 | # Add words to the vocabulary. 102 | for i, word in enumerate(words): 103 | vocab.add_word(word) 104 | vocab.add_word('') 105 | 106 | return vocab 107 | 108 | 109 | def main(data_path, data_name): 110 | vocab = build_vocab(data_path, data_name, caption_file=annotations, threshold=5) 111 | serialize_vocab(vocab, 'vocab/%s_vocab.json' % data_name) 112 | print("Saved vocabulary file to ", 'vocab/%s_vocab.json' %(data_name)) 113 | 114 | 115 | 116 | if __name__ == '__main__': 117 | parser = argparse.ArgumentParser() 118 | parser.add_argument('--data_path', default='data') 119 | parser.add_argument('--data_name', default='sydney_precomp', 120 | help='{coco,f30k}') 121 | opt = parser.parse_args() 122 | main(opt.data_path, opt.data_name) 123 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | absl-py==0.11.0 2 | argon2-cffi==20.1.0 3 | asn1crypto==0.24.0 4 | astunparse==1.6.3 5 | async-generator==1.10 6 | attrs==20.3.0 7 | backcall==0.2.0 8 | bleach==3.3.0 9 | cachetools==4.2.0 10 | certifi==2020.12.5 11 | cfbclient==1.1.0 12 | cffi==1.14.4 13 | chardet==4.0.0 14 | click==8.0.4 15 | cryptography==3.3.2 16 | cycler==0.10.0 17 | dataclasses==0.8 18 | decorator==4.4.2 19 | defusedxml==0.6.0 20 | diplomat==1.4.0 21 | entrypoints==0.3 22 | faiss==1.5.3 23 | filelock==3.4.1 24 | flatbuffers==1.12 25 | future==0.18.2 26 | gast==0.3.3 27 | google-auth==1.24.0 28 | google-auth-oauthlib==0.4.2 29 | google-pasta==0.2.0 30 | gr-record==1.12 31 | grpcio==1.32.0 32 | gviz-api==1.9.0 33 | h5py==2.10.0 34 | huggingface-hub==0.0.12 35 | idna==2.6 36 | importlib-metadata==3.4.0 37 | importlib-resources==5.4.0 38 | ipykernel==5.1.1 39 | ipython==7.16.3 40 | ipython-genutils==0.2.0 41 | ipywidgets==7.6.3 42 | jedi==0.17.2 43 | Jinja2==2.11.3 44 | joblib==1.0.1 45 | jsonschema==3.2.0 46 | jupyter==1.0.0 47 | jupyter-client==6.1.11 48 | jupyter-console==6.2.0 49 | jupyter-core==4.7.0 50 | jupyter-http-over-ws==0.0.8 51 | jupyterlab-pygments==0.1.2 52 | jupyterlab-widgets==1.0.0 53 | Keras-Preprocessing==1.1.2 54 | keyring==10.6.0 55 | keyrings.alt==3.0 56 | kiwisolver==1.3.1 57 | lightgbm==3.2.1 58 | llvmlite==0.36.0 59 | Markdown==3.3.3 60 | MarkupSafe==1.1.1 61 | matplotlib==3.3.3 62 | missingno==0.5.0 63 | mistune==0.8.4 64 | nbclient==0.5.1 65 | nbconvert==6.0.7 66 | nbformat==4.4.0 67 | nest-asyncio==1.4.3 68 | nltk==3.6.7 69 | notebook==6.4.10 70 | numba==0.53.1 71 | numpy==1.21 72 | oauthlib==3.1.0 73 | opencv-contrib-python-headless==4.5.5.64 74 | opt-einsum==3.3.0 75 | packaging==20.8 76 | pandas==1.1.5 77 | pandocfilters==1.4.3 78 | parso==0.7.1 79 | pexpect==4.8.0 80 | pickleshare==0.7.5 81 | Pillow==9.0.1 82 | plotly==5.1.0 83 | prometheus-client==0.9.0 84 | prompt-toolkit==3.0.13 85 | protobuf==3.15.0 86 | pstatsd==1.2.3 87 | ptyprocess==0.7.0 88 | pyasn1==0.4.8 89 | pyasn1-modules==0.2.8 90 | pycparser==2.20 91 | pycrypto==2.6.1 92 | Pygments==2.7.4 93 | pygobject==3.26.1 94 | pyparsing==2.4.7 95 | pyrsistent==0.17.3 96 | python-apt==1.6.5+ubuntu0.5 97 | python-consul==1.1.0 98 | python-dateutil==2.8.1 99 | pytz==2021.1 100 | pyxdg==0.26 101 | PyYAML==6.0 102 | pyzmq==21.0.1 103 | qtconsole==5.0.2 104 | QtPy==1.9.0 105 | redis==3.5.3 106 | regex==2022.3.15 107 | requests==2.25.1 108 | requests-oauthlib==1.3.0 109 | rsa==4.7 110 | ruamel.yaml==0.17.21 111 | ruamel.yaml.clib==0.2.6 112 | sacremoses==0.0.49 113 | scikit-learn==0.24.2 114 | scikit-image 115 | scipy==1.5.4 116 | seaborn==0.11.1 117 | SecretStorage==2.3.1 118 | Send2Trash==1.5.0 119 | six==1.15.0 120 | skipthoughts==0.0.1 121 | ssh-import-id==5.7 122 | tenacity==8.0.1 123 | tensorboard==2.4.1 124 | tensorboard-logger==0.1.0 125 | tensorboard-plugin-profile==2.5.0 126 | tensorboard-plugin-wit==1.8.0 127 | termcolor==1.1.0 128 | terminado==0.9.2 129 | testpath==0.4.4 130 | threadpoolctl==2.2.0 131 | timm==0.5.4 132 | tlcache==0.3.1 133 | tokenizers==0.10.3 134 | torch==1.8.1 135 | torchvision==0.9.1 136 | tornado==6.1 137 | tqdm==4.63.1 138 | traitlets==4.3.3 139 | transformers==4.8.1 140 | typing-extensions==3.7.4.3 141 | urllib3==1.26.5 142 | wcwidth==0.2.5 143 | webencodings==0.5.1 144 | Werkzeug==1.0.1 145 | widgetsnbextension==3.5.1 146 | wrapt==1.12.1 147 | xgboost==1.4.2 148 | zipp==3.4.0 149 | -------------------------------------------------------------------------------- /test_data/imgs/README.md: -------------------------------------------------------------------------------- 1 | Download SLM_testimgs.zip, unzip it and put it here. 2 | 3 | ```angular2 4 | -SLM 5 | ---... 6 | ---test_data 7 | -----... 8 | -----imgs 9 | -------0.jpg 10 | -------1.jpg 11 | -------2.jpg 12 | -------... 13 | -------21.jpg 14 | ``` -------------------------------------------------------------------------------- /test_data/visual_testdata.py: -------------------------------------------------------------------------------- 1 | # ** 2 | # * Copyright @2022 AI, AIRCAS. (mails.ucas.ac.cn) 3 | # 4 | # @author yuanzhiqiang 5 | # 2022/04/03 6 | 7 | import json 8 | import os 9 | 10 | import matplotlib.image as imgplt 11 | import matplotlib.pyplot as plt 12 | 13 | 14 | def analyze_samples(json_path): 15 | # load json 16 | with open(json_path,'r',encoding='utf8')as fp: 17 | json_data = json.load(fp) 18 | 19 | # analyze 20 | print("=========================") 21 | print("Lens of items: {}\n".format(len(json_data))) 22 | print("===== Map Relation ======") 23 | for idx, item in enumerate(json_data): 24 | print("Idx:{}, Filename:{}, Caption:{}".format(idx, json_data[idx]['jpg_name'], json_data[idx]['caption'].replace("\n", ""))) 25 | 26 | return json_data 27 | 28 | def visual_data(png_path, json_data, show_idx): 29 | 30 | # load show data 31 | show_json_data = json_data[show_idx] 32 | finalname = os.path.join(png_path, show_json_data['jpg_name']) 33 | plotpoints = show_json_data['points'] 34 | captiondata = show_json_data['caption'] 35 | 36 | print("\n=========================") 37 | print("===== Visual Data=========") 38 | print("Idx: {}".format(show_idx)) 39 | print("Filename: {}".format(finalname)) 40 | print("Caption: {}".format(captiondata.replace("\n", ""))) 41 | print("Annotations: {}".format(plotpoints)) 42 | 43 | # visual data 44 | pic = imgplt.imread(finalname) 45 | plt.imshow(pic) 46 | plt.title(show_json_data['jpg_name']) 47 | 48 | xdata = [] 49 | ydata = [] 50 | for k in range(len(plotpoints)): 51 | xdata.clear() 52 | ydata.clear() 53 | for j in range(len(plotpoints[k])): 54 | item = plotpoints[k][j] 55 | x = item[0] 56 | y = item[1] 57 | plt.scatter(x, y, s=25, c='r') 58 | xdata.append(x) 59 | ydata.append(y) 60 | plt.plot(xdata, ydata, c='b') 61 | plt.plot([xdata[0], xdata[j]], [ydata[0], ydata[j]], c='b') 62 | plt.show() 63 | 64 | 65 | if __name__ == "__main__": 66 | png_path = "./imgs/" 67 | json_path = "./annotations/anno.json" 68 | show_idx = 44 69 | 70 | # analyze samples 71 | json_data = analyze_samples(json_path) 72 | 73 | # plot one sample 74 | visual_data(png_path, json_data, show_idx) 75 | 76 | --------------------------------------------------------------------------------