├── README.md
├── grids.txt
├── predict_geolocation.ipynb
└── predict_places365.ipynb


/README.md:
--------------------------------------------------------------------------------
 1 | # Geo-location Tutorial
 2 | 
 3 | Download model:
 4 | https://s3.amazonaws.com/mmcommons-tutorial/models/RN101-5k500-0012.params
 5 | https://s3.amazonaws.com/mmcommons-tutorial/models/RN101-5k500-symbol.json
 6 | 
 7 | Geolocation model inspired by ideas presented in:
 8 | PlaNet - Photo Geolocation with Convolutional Neural Networks (ECCV 2016),
 9 | Tobias Weyand, Ilya Kostrikov, James Philbin 
10 | https://research.google.com/pubs/pub45488.html
11 | 
12 | ## Data and Classes
13 | Our data come from the geotagged images in the YFCC100M Multimedia Commons dataset.
14 | Training, validation, and test images are split so that images uploaded by the same person do not appear in multiple sets.
15 | Classes are created with the training data using [Google's S2 Geometry Library](https://code.google.com/archive/p/s2-geometry-library/) 
16 | as described in the PlaNet paper above. The classes are defined in `grids.txt` where the i-th line is the i-th class and the columns are: 
17 | `S2 Cell Token, Latitude, Longitude`.
18 | 
19 | ## Difference between our model and PlaNet:
20 | 
21 | |               | Ours  | PlaNet |
22 | |---------------|-------|--------|
23 | |Dataset source        |Multimedia Commons|Images crawled from the web|
24 | |Training set|33.9 million|91 million|
25 | |Validation|1.8 million|34 million|
26 | |S2 Cell Partitioning|t_1=5000, t_2=500 ==> 15,527 cells|t_1=10,000, t_2=50 ==> 26,263 cells|
27 | |Model|ResNet-101|GoogleNet|Inception|
28 | |Optimization|SGD with Momentum and LR Schedule|Adagrad|
29 | |Training time| 9 days on 16 NVIDIA K80 GPUs (p2.16xlarge), 12 epochs|2.5 months on 200 CPU cores|
30 | |Framework|MXNet | DistBelief|
31 | |Test set|Placing Task 2016 Test Set (1.5 million Flickr images)|2.3 M geo-tagged Flickr images|
32 | 
33 | ## Result
34 | ### Im2GPS test set 
35 | The values indicate the percentages of images within test set that were correctly localized within the given distance. 
36 | 
37 | |Method|1km|25km|200km|750km|2500km|
38 | |------|---|----|-----|-----|------|
39 | |PlaNet|8.4%|24.5%|37.6%|53.6%|71.3%|
40 | |Ours|16.8%|39.2%|48.9%|67.9%|82.2%|
41 | 
42 | ### Flickr Images 
43 | Note that these result in the table are not directly comparable as the test set images used in PlaNet is not publicly released. 
44 | The values indicate the percentages of images within test set that were correctly localized within the given distance. 
45 | 
46 | |Method|1km|25km|200km|750km|2500km|
47 | |------|---|----|-----|-----|------|
48 | |PlaNet|3.6%|10.1%|16.0%|28.4%|48.0%|
49 | |Ours|6.2%|13.5%|20.8%|35.6%|55.2%|
50 | 
51 | 
52 | 


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