VID-Trans-ReID: Enhanced Video Transformers for Person Re-identification

원문 초록

Video-based person Re-identification (Re-ID) has received increasing attention recently due to its important role within surveillance video analysis. Video-based Re- ID expands upon earlier image-based methods by extracting person features temporally across multiple video image frames. The key challenge within person Re-ID is extracting a robust feature representation that is invariant to the challenges of pose and illumination variation across multiple camera viewpoints. Whilst most contemporary methods use a CNN based methodology, recent advances in vision transformer (ViT) architectures boost fine-grained feature discrimination via the use of both multi-head attention without any loss of feature robustness. To specifically enable ViT architectures to effectively address the challenges of video person Re-ID, we propose two novel modules constructs, Temporal Clip Shift and Shuffled (TCSS) and Video Patch Part Feature (VPPF), that boost the robustness of the resultant Re-ID feature representation. Furthermore, we combine our proposed approach with current best practices spanning both image and video based Re-ID including camera view embedding. Our proposed approach outperforms existing state-of-the-art work on the MARS, PRID2011, and iLIDS-VID Re-ID benchmark datasets achieving 96.36%, 96.63%, 94.67% rank-1 accuracy respectively and achieving 90.25% mAP on MARS.