Rethinking Triplet Loss for Domain Adaptation

The gap in data distribution motivates domain adaptation research. In this area, image classification intrinsically requires the source and target features to be co-located if they are of same class. However, many works only take a global view gap. That is, make distributions globally overlap; does not necessarily lead feature co-location at class level. To resolve problem, we study metric learning context adaptation. Specifically, introduce similarity guided constraint (SGC). implementation, SGC takes form triplet loss. loss is integrated into network as an additional objective term. Here, consists two images another different Albeit simple, working mechanism our method interesting insightful. Importantly, triplets sampled from domains. From micro perspective, by enforcing on every possible triplet, domains but mapped nearby, those classes far apart. macro ensures that cross-domain similarities preserved, leading intra-class compactness inter-class separability. Extensive experiment four datasets shows yields significant improvement over baselines has competitive accuracy with state-of-the-art results.