Interventional Supervised Learning for Person Re-identification

Person re-identification (ReID) is a problem of retrieving pedestrian images. The complexity of ReID is attributed to image resolution, shooting angles, and object occlusion. With the evolution of computer vision, researchers have proposed effective solutions to tackle ReID. Causal inference has demonstrated its utility across various computer vision tasks. This paper investigates the potential of causal inference technology to enhance ReID performance. We construct a structural causal model for ReID and determine that dataset knowledge can constrain models from grasping genuine causality by affecting image features. Consequently, we propose intervention-based methods for supervised learning ReID (ISReID), introducing three effective backdoor adjustment schemes to intervene in supervised learning ReID models. To evaluate our approach's effectiveness, we conduct experiments on four widely-used ReID datasets: DukeMTMC-reID, Market-1501, CUHK03, and MSMT17. The experimental results confirm the efficacy of our backdoor adjustment schemes in enhancing supervised learning ReID models' performance. The source code can be found at https://github.com/SCNU203/ISReID.