Generalized Contrastive Loss preprint is out!

I’m happy to share the preprint of my latest paper, titled “Generalized Contrastive Optimization of Siamese Networks for Place Recognition”, which can be found on arXiv.

I propose a new contrastive loss function, namely the Generalized Contrastive Loss, that relies on a graded definition of image similarity ground truth. This function, together with the new ground truth allows us to train a siamese architecture without relying on overcomplicated pair mining strategies and using only simple pooling layers (GeM and average), outperforming more complex state-of-the-art approaches like NetVLAD.

Find the testing code and new sets of labels for the MSLS, TB-Places, and 7Scenes datasets on the github repository.