Revisiting Im2GPS

December 19, 2018 · View on GitHub

Code & data for our "Revisiting IM2GPS in the Deep Learning Era" paper

PyTorch version: to be released

Prerequisites:

Test on new images:

Download reference features from http://www.mediafire.com/file/7l2dqi7f7obzu57/reference_im2gps.zip and unzip the .mat files to "reference" folder
Download caffe model from http://www.mediafire.com/file/d9hifat6qi4tu2m/7000only__iter_167400.caffemodel
Run main.m to predict gps coordinate of an example image

Test on im2gps-test-set:

Download the test set from http://graphics.cs.cmu.edu/projects/im2gps and unzip the images to "query" folder
Run main_im2gps_test.m
The numbers might not be exactly the same as reported in the paper because of FLANN randomization or image decoding.

Datasets:

Im2GPS3k, used for testing in our paper: this is part of the original Im2GPS dataset, here's the 3000 (full resolution) images http://www.mediafire.com/file/7ht7sn78q27o9we/im2gps3ktest.zip
Im2GPS, used for training/referencing: more than 5 million images, this is part of the original Im2GPS dataset. We are not able to provide access to the images, though you can download the reference features (above), look up the file names and infer the original sources on Flickr.
YFCC100M, used for referencing: due to time constraint, we were only able to download ~22 million images for the experiments reported in the paper. The full dataset can be found here https://webscope.sandbox.yahoo.com/catalog.php?datatype=i&did=67
YFCC4k, used for testing in our paper: actually 4536 images. We've lost our YFCC100M data, but still have the list of image-id here: http://www.mediafire.com/file/8v2j565997i5jed/0aaaa.r.imagedata.txt, from which you can look up the sources/URLs in the YFCC100M dataset. Update: we got the images: www.mediafire.com/file/3og8y3o6c9de3ye/yfcc4k.zip

Reference:

Nam Vo, Nathan Jacobs and James Hays. "Revisiting IM2GPS in the Deep Learning Era". ICCV 2017.