LoOp: Looking for Optimal Hard Negative Embeddings for Deep Metric Learning

Deep metric learning has been effectively used to learn distance metrics for\ndifferent visual tasks like image retrieval, clustering, etc. In order to aid\nthe training process, existing methods either use a hard mining strategy to\nextract the most informative samples or seek to generate hard synthetics using\nan additional network. Such approaches face different challenges and can lead\nto biased embeddings in the former case, and (i) harder optimization (ii)\nslower training speed (iii) higher model complexity in the latter case. In\norder to overcome these challenges, we propose a novel approach that looks for\noptimal hard negatives (LoOp) in the embedding space, taking full advantage of\neach tuple by calculating the minimum distance between a pair of positives and\na pair of negatives. Unlike mining-based methods, our approach considers the\nentire space between pairs of embeddings to calculate the optimal hard\nnegatives. Extensive experiments combining our approach and representative\nmetric learning losses reveal a significant boost in performance on three\nbenchmark datasets.\n