Occlusion robust re-identification

Abstract

For efficiently exploiting the bulk of information obtained from city-wide CCTV systems, object (pedestrians, vehicles) tracking and re-identification across multiple cameras are essential tasks. Object re-identification (ReID) has extensive utility for analyzing and accurately merging objects' trajectories across multiple non-overlapping cameras. Despite being a very fertile ground for research which attracted growing attention in the past decades, re-identification is still considered a very challenging problem. Moreover, most available datasets contain unoccluded target images and thus the networks trained on them fail to generalize to practical examples where occlusions and various artifacts are common. In this work, we propose an adversarial training strategy to robustify re-identification for occluded person and vehicle images in intelligent surveillance systems. Our method adversarially learns to generate occlusion masks that diminish the re-identification network performance, thus challenging its adversary, the re-identification network, to enhance re-ID performance against physical occluders. Extensive experiments performed on public datasets as well as a personally collected dataset verify the effectiveness of the proposed approach.