Multi-Pose Learning based Head-Shoulder Re-identification

The whole body of person is probably invisible in video surveillance because of occlusion and view angles (such as in crowded public places), on which occasion conventional person re-identification (i.e., whole-body based Re-ID) approaches may not work. To address this problem, we propose a novel deep pairwise model based on multi-pose learning (MPL) which aims at head-shoulder part instead of the whole body. The proposed method explicitly tackles pose variations by learning an ensemble verification conditional probability distribution about relationship among multiple poses. To facilitate the research on this problem, we contribute three head-shoulder datasets based on CUHK03, CUHK01 and VIPeR. Experiments on these datasets demonstrate that our proposed method achieves the state-of-the-art performance.