Pose-Guided Multi-Granularity Feature Learning for Occluded Person Re-Identification

Abstract

Persons are often occluded in real-world applications of person re-identification.To alleviate the occlusion problem, this paper proposes a pose-guided multi-granularity feature learning method for occluded person re-identification.At first, the residual atrous spatial pyramid pooling module is used to expand the receptive field to extract more hierarchical pedestrian features.Next, the visible head-and-shoulder region feature and the underneath region feature from the heatmap extracted by the pedestrian estimation algorithm are calculated.Finally, the multi-granularity strategy is adopted to learn the pedestrian features of different hierarchies of information in the visible pedestrian region.Experimental results on the Occluded-DukeMTMC, Occluded-REID, and Market1501 datasets demonstrate the effectiveness of our proposed method.