TriEP: Expansion-Pool TriHard Loss for Person Re-Identification

Person re-identification aims to identify the same person across different cameras, which is widely applied in the intelligent monitoring field. The research of TriHard loss has been verified to improve performance efficiently in the person re-identification task. However, TriHard loss considers the hardest positive sample and the hardest negative sample exclusively, which ignores the remaining samples. To stuff this gap, we propose an expansion-pool TriHard (TriEP) loss which can give attention to the hardest samples and other samples. Initially, the equal-label sample pools are expanded based on the labels of the hardest samples. Then, the statistics of the sample pool are calculated according to the distribution of samples. Finally, the dynamic penalty is imposed on TriHard loss to construct TriEP loss. Extensive experiments on Market-1501, DukeMTMC, and occluded datasets prove the superiority of the proposed TriEP loss. Compared with the baseline model, a noticeable performance improvement can be obtained after embedding the proposed TriEP loss.