Selecting Effective Triplet Contrastive Loss for Domain Alignment

Deep Neural Networks (DNNs) have achieved great success in many applications. However, the performance of DNNs degrades significantly in unseen scenarios. Triplet contrastive loss as a domain alignment method has been used in domain generalization to solve above problems. However, there is still no method to guide how to select an effective triplet contrastive loss. In this paper, we derive three triplet contrastive losses as upper bounds for contrastive loss. We find an effective triplet contrastive loss by comparing the range of the three triplet contrastive losses and the rate of hard negative loss pairs. Then, the triplet contrastive loss is applied in domain alignment tasks to explore class invariant representations, demonstrating its effectiveness on five standard benchmarks.