Exhaustive hard triplet mining loss for Person Re-Identification

Abstract

Person reidentification (Re-ID) is an important task in computer vision and has many applications in video-based surveillance. Recently, the triplet loss has been popular in the deep learning framework for person Re-ID. It is particularly important to note that the selection of hard triplets has significant influence on the performance of the learned deep model. However, the existing triplet losses only focus on some specific forms of hard triplets, thus leading to weaker generalization capability. To address this issue, we propose a novel variant of the triplet loss, named exhaustive hard triplet mining loss (EHTM), which is able to deal with various forms of hard triplets in a comprehensive manner. Moreover, the proposed loss comprises a term to facilitate distinguishing different identities by directly narrowing intraclass distances and indirectly enlarging interclass distances. We also provide an effective training strategy to further enhance model performance. Extensive experiments on several benchmark datasets show that our method outperforms state-of-the-art approaches by a large margin.