Learning deep part-aware embedding for person retrieval

Person retrieval is an important vision task, aiming at matching the images of same person under various camera views. The key challenge lies in large intra-class variations among images. Therefore, how to learn discriminative feature representations becomes core problem. In this paper, we propose a deep part-aware representation learning method for retrieval. First, improved triplet loss introduced such that global from identity are closely clustered. Meanwhile, localization branch proposed automatically localize those person-wise parts or regions, only using labels weakly supervised manner. Via simultaneously guided by and branch, can further improve performance Through extensive set ablation studies, verify each contributes boosts method. Our model obtains superior (or comparable) compared state-of-the-art methods on four public datasets. On CUHK03-labeled dataset, instance, increases 73.0% mAP 77.9% rank-1 accuracy 80.8% (+7.8%) 83.9% (+6.0%) accuracy.