MGD: Mask Guided De-occlusion Framework for Occluded Person Re-identification

摘要

Person re-identification (ReID) is a challenging task in computer vision area due to the dramatic changes across different non-overlapping camera views, e.g., lighting, view angle, and pose, among which occlusion is one of the hardest challenges. Recently, occluded person re-identification (Occluded-ReID) is proposed to address this problem. Nevertheless, current occluded-ReID methods focus on how to learn a matching function between partial-body images and full-body images while ignore the structural information of the full body. To handle this problem, we propose a novel framework called Mask Guided De-occlusion (MGD) for occluded Person Re-identification. The MGD mainly consists of three components, i.e., a Coarse-to-Fine Mask Generation (CFMG) module, a Person De-Occlusion (PDO) module and a Person Feature Extractor (PFE). The key module CFMG aims to locate the occlusion areas by manipulating the instance segmentation masks through a two-stage process. The proposed PDO module is to reconstruct the occluded pedestrian. After that, all the images are fed into the PFE module to obtain their feature vectors. With PDO and CFMG modules, the proposed method MGD reduces the impact of occlusions and thus improves the performance of Occluded-ReID. The extensive experiments conducted on several public occluded ReID datasets show that our method is effective and outperforms the state-of-the-art methods.