Reasoning and Tuning: Graph Attention Network for Occluded Person Re-Identification

Abstract

Occluded person re-identification (re-id) aims to match occluded person images to holistic ones. Most existing works focus on matching collective-visible body parts by discarding the occluded parts. However, only preserving the collective-visible body parts causes great semantic loss for occluded images, decreasing the confidence of feature matching. On the other hand, we observe that the holistic images can provide the missing semantic information for occluded images of the same identity. Thus, compensating the occluded image with its holistic counterpart has the potential for alleviating the above limitation. In this paper, we propose a novel Reasoning and Tuning Graph Attention Network (RTGAT), which learns complete person representations of occluded images by jointly reasoning the visibility of body parts and compensating the occluded parts for the semantic loss. Specifically, we self-mine the semantic correlation between part features and the global feature to reason the visibility scores of body parts. Then we introduce the visibility scores as the graph attention, which guides Graph Convolutional Network (GCN) to fuzzily suppress the noise of occluded part features and propagate the missing semantic information from the holistic image to the occluded image. We finally learn complete person representations of occluded images for effective feature matching. Experimental results on occluded benchmarks demonstrate the superiority of our method.