Pooling body parts on feature maps for misalignment robust person re-identification
Yuting Liu, Qijun Zhao, Zhihong Wu · 2018
Misalignment caused by poorly detected bounding boxes or varying poses of human bodies is a critical challenge to robust person re-identification (re-ID) systems. Most previous works extract features either from the detected entire pedestrian images or from local image patches that are globally aligned or cropped according to human body landmarks. However, these methods still suffer from misalignment ofhuman bodies in different images. In this paper, we present a Deep Joint Learning (DJL) network to fulfill misalignment robust person re-ID. It locally aligns the human bodies by pooling the features around the body parts on feature maps, and jointly optimizes the global and aligned local features to further enhance the discriminative capability oflearned feature representations. Experimental results on Market-1501 and CUHK03 datasets show that our method can effectively handle the misalignment induced intra-class variations and yield competitive accuracy particularly on poorly aligned pedestrian images.