Overcoming Occlusion in Person Re-Identification: A Multi-Level Attention Transformer Approach

Abstract

Person re-identification (ReID) in real world surveillance scenarios is a very challenging problem, in which occlusions are a major culprit that can severely degrade the performance of existing systems. In this paper, we proceed one step closer towards solving this critical problem by proposing a novel Multi Level Attention Mechanism (MLAM) for occluded person re identification. Our approach uses spatial, channel and global context attention to tackle different occlusion cases from partial to severe.

The proposed method integrates two key architectures: the Occlusion-Aware ReID Transformer (OART) and the Multi-Level Attention Transformer Network (MLATN). Specifically, we demonstrate that the proposed framework enables adaptive feature extraction and occlusion aware fusion, which brings large robustness gains when used for adaptive ReID in real world challenging environments.

Study evaluate the approach through extensive experiments on the challenging datasets, Occluded-DukeMTMC and OccludedREID, and demonstrate the superiority of our approach. For the Occluded DukeMTMC, the MLAM achieves state of the art performance with 2.7% and 5.1% Rank 1 accuracy and mean Average Precision (mAP) respectively. We also propose the Occlusion Robustness Index (ORI): We present a new model invariant metric to quantify model resilience to occlusions.

Beyond surveillance, the results of this research are applicable to autonomous driving, robotics, and augmented reality. Nevertheless, significant advances have been made, which casts into sharp relief significant ethical issues around privacy and protection of information, and a need for accountable development and deployment of such technologies. Towards this end, we believe this work presents a large step towards occluded person reidentification and the development of robust adaptable vision recognition systems for difficult real world circumstances.