ECCV 2020
Raul Gomez, Jaume Gibert, Lluis Gomez and Dimosthenis Karatzas

Paper
Blog Post

This code needs intense cleaning and simplification that will be done as soon as possible

LocSens is able to retrieve images related to a query word and near to a given location:

LocSens is able to tag images exploiting both visual and location information of an image, outperforming location agnostic tagging methods:

Here we explain roughly how this first version of the code works. Each script includes a short explanation of what it does.

Contains the code to generate the hashtag vocabulary, the splits and the query sets. It works over the YFCC100M dataset, which is available here.

Contains the exact splits we used in our paper.

Contain the code to train this models as explained in the paper chapter "Learning with Hashtag Supervision."

Contains the code to train LocSens.

Contains the code to evaluate both the location agnostic models and LocSens. The code the compute the embeddings used in this evaluations scripts is located in each of the models folders.