Deep Metric Learning with Online Hard and Soft Selection for Person Re-identification

Deep metric learning has been widely used for image retrieval and verification tasks. Traditional contrastive loss and triplet loss depend highly on the selection of pair/triplet images. It makes the training process unstable and uncomplete. In this paper, we propose a novel global level loss function that considers histograms for intra distances within class and inter distances between different classes. We compared two forms of global level loss (hard selection based loss and soft selection based loss) and both achieved better result than traditional triplet loss, multi class N pair loss and other related works. The experiment is conducted on the person re-identification dataset Market 1501 and DukeMTMC-reID.