2 code implementations • ECCV 2020 • Djebril Mekhazni, Amran Bhuiyan, George Ekladious, Eric Granger
We argue that for pair-wise matchers that rely on metric learning, e. g., Siamese networks for person ReID, the unsupervised domain adaptation (UDA) objective should consist in aligning pair-wise dissimilarity between domains, rather than aligning feature representations.
no code implementations • 11 Feb 2020 • George Ekladious, Hugo Lemoine, Eric Granger, Kaveh Kamali, Salim Moudache
To this end, a dual-triplet loss is introduced for metric learning, where two triplets are constructed using video data from a source camera, and a new target camera.