Region selection for occluded person re-identification via policy gradient

Person re-identification (re-id) has attracted lots of attention in the past few years. However, most of work in this area focus on the re-identification of the holistic pedestrian images. In the real application scenarios, the task of re-identification is usually influenced by the occlusion problems. Most of previous work tackle occlusion problem through pose estimation, human parsing or manually labelling occlusion objects. Such reliance on additional an- notation information severely limits the generalization ability on practical usage. To address this problem, we propose a novel region selection learning strategy based on the policy gradient to remove irrelevant parts with- out using any extra information outside dataset. A transformer-based feature extractor is also constructed to learn discriminate features with self-attention mechanism. We evaluate the performance of the proposed method on three occluded re-identification datasets. The experiments show that we achieve 87%, 85.2% and 68.3% in Rank-1 accuracy on Occluded-REID, Partial-REID and Occluded-Duke datasets respectively.(c) 2023 Elsevier B.V. All rights reserved.