Triplet online instance matching loss for person re-identification

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