Release corresponding to AAAI 2020 version

Images

We use 750k images downloaded from Flick, where each image has a geotag to specify the ground-truth location. We randomly split them into training and test sets, of size 90% and 10% of total images respectively. In order to extend to our work to the black-box setting, we further evenly and exclusively split the training data into two parts. We discretize the globe into 512 cells and label each image based on its latitude and longitude.

You can find model1.txt,model2.txt and text.txt under image_links. You will find the URLs and labels for the images we used from these files.

Image Features

We provide the trained weights under weights folder. weights/model1/iter* is the weighted trained on half of the training data. weights/model2/iter* is the weights trained on the other half, for the sake of applying for the black-box protection. You can use these weights to get the image features for the photos from your own photo albums.

Since it's impossible to provide the raw images dataset, we provide the extracted feature for each image in the test set, which is saved under imgs_features folder. You may use them to generate the score matrices for the synthetic albums we created.

Synthetic Albums

You can find the album information under the albumlists folder. For each .txt file, there are three columns for each row, column 1 : album ID, column 2 : label, column 3: photos in this album (local_address/image_name).

Flickr Albums

We provide the Flickr album score matrices directly. They are saved under Flickr_album.