Part-aware network: a simple but efficient method for occluded person re-identification

Abstract

Person re-identification is to query person across cameras and occlusion is one of the difficulties. Previous works have proved that local feature extraction and alignment are critical for occluded person re-identification. However, directly horizontal partition causes mis-alignment and extra-cue methods highly depend on the quality of extra-cues. In this work, we propose a novel architecture including a weakly supervised mask generator without introducing extra-cues to create fine-grained semantic masks for local feature extraction and alignment, and a weight-shared fully connection to control the balance of local and global features. We also propose a general form of weighted pooling to improve gradient transfer, which gets rid of the probability explanation with softmax. Moreover, we unravel that there is a conflict between local branches and global branch, and a buffer convolution layer helps to fix this conflict. Extensive experiments show the effectiveness of our proposed method on occluded and holistic ReID tasks. Specially, we achieve 62.5% Rank-1 and 52.6% mAP (mean average precision) scores on the occluded-duke dataset.