Multi-level and multi-scale horizontal pooling network for person re-identification

Abstract

Person re-identification (Re-ID) is the task of matching a target person across different cameras, which has drawn extensive attention in computer vision and has become an essential component in the video surveillance system. Despite recent remarkable progress, person re-identification methods are either subject to the power of feature representation, or give equal importance to all examples. To mitigate these issues, we introduce a simple, yet effective, Multi-level and Multi-scale Horizontal Pooling Network (MMHPN) for person re-identification. Concretely, our contributions are three-fold:1) we take partial feature representation into account at different pooling scales and different semantic levels so that various partial information is obtained to form a robust descriptor; 2) we introduce a Part Sensitive Loss (PSL) to reduce the effect of easily classified partition to facilitate training of the person re-identification network, 3) we conduct extensive experimental results using the Market-1501, DukeMTMC-reID and CUHK03 datasets and achieve mAP scores of 83.4%, 75.1% and 65.4% respectively on these challenging datasets.