Part Matching with Multi-Level Attention for Person Re-Identification

Person re-identification, which aims to match two instances across different cameras still remains challenging. One of the biggest obstacles is the misalignment problem, namely instances cannot be spatially aligned due to the large variations in scale, pose, etc. In this work, we proposed a Multi-level Attention Network associated with a Part Matching Distance to address this issue. Specially, the Multi-level Attention Network consists of both pixel-level and part-level attention, which together form a coarse-to-fine flow to extract discriminative features of each instance. The Part Matching Distance formulates the distance between two instances with a part searching scheme under the constrains of the spatial structures, which can further tackle the large-scale variations at the inference stage. Both quantity and quality results on three popular person Re-ID benchmarks and one person search benchmark show that our method outperforms all previous works by a notable margin.