Home

Awesome

GT-CrossView

This contains the data and model for training crossview image ranking method described in: Localizing and Orienting Street Views Using Overhead Imagery, ECCV 2016. https://lugiavn.github.io/gatech/crossview_eccv2016/

Data

Download from https://www.mediafire.com/folder/f4gga3h86d659/GTCrossView and unzip them into gt-crossview/data/

Caffe model

You will need to install caffe with our proposed DBL log loss function: https://github.com/lugiavn/caffe/tree/embedding_losses

Note

Note that both data and the model is not the exact same one we used in the paper.

For data, we used the original high resolution images and resized them, making the entire dataset smaller so that it's easy to download and experiment with. For the caffe model, we reimplemented/refactored the original code so that it's clean. Therefore the number might be different from the paper; if you train and test the model on the data here for comparison, the result would be:

TablesDenver test setDetroit test setSeattle test set
Best classification
1 rotation crop93.089.589.4
4 rotation crops93.489.989.8
16 rotation crops93.790.189.9
Recall at 0.01
1 rotation crop58.252.046.1
4 rotation crops66.859.853.3
16 rotation crops69.561.355.5

or you can use a model that we trained: https://github.com/lugiavn/gt-crossview/issues/1