Triplet online instance matching loss for person re-identification

Abstract

Mining the shared features of the same identity in different scenes and the unique features of different identities in the same scene are the most significant challenges in the field of person re-identification (ReID). The Online Instance Matching (OIM) loss function and triplet loss function are the main methods for person ReID. Unfortunately, both of them have drawbacks. The OIM loss treats all samples equally and puts no emphasis on hard samples. The triplet loss processes batch construction in a complicated and fussy way and converges slowly. For these problems, we propose a Triplet Online Instance Matching (TOIM) loss function, which emphasizes hard samples and improves the person ReID accuracy effectively. It combines the advantages of the OIM loss and triplet loss and simplifies the batch construction process, which leads to quicker convergence. It can be trained on-line when handling the joint detection and identification task. To validate our loss function, we collect and annotate a large-scale benchmark dataset (UESTC-PR), which contains 499 identities and 60,437 images taken from surveillance cameras. We evaluated our proposed loss function on the Duke, Marker-1501 and UESTC-PR datasets using ResNet-50, and the results show that our proposed loss function outperforms the baseline methods, including Softmax loss, OIM loss and triplet loss.