Pedestrian-Aligned Multiscale Features Network for Person Re-identification

Person re-identification intends to seek out and match a target pedestrian from different cameras. In practice, images captured by different cameras typically have variances in complex scenes. In addition, large-scale datasets generally adopt automatic detectors to obtain pedestrian detection boxes, and detection errors can lead to image misalignment problems. Both complex scenes and detection errors reduce the robustness of pedestrian characteristics and exasperate the capability of pedestrian alignment. To tackle the above conundrums, we design a novel end-to-end method called Pedestrian-Aligned Multiscale Features Network (PA-MFN), integrating pedestrian alignment and multiscale features network through spatial transformer networks (STN). We employ pedestrian alignment network to work out image misalignment problems satisfactorily. Additionally, we carefully design multiscale features network to extract both global and local features. Comprehensive experiments indicate that our approach gains superior performance on Market-1501, DukeMTMC-reID as well as CUHK03 datasets.