Top-View Person Re-Identification

Although considerable progress has been made in person Re-IDentification (ReID), challenging problems remain. With the growing importance of personal privacy and security, a new task called top-view person ReID is introduced, which contains particularly challenging viewpoint variations and limited discriminative information. Compared to general person ReID, we can obtain the capture angle of a top-view person image through point cloud and head-shoulder detection from deployed binocular cameras. Existing person ReID methods mainly focus on learning view-invariant feature representations, but they are not effective for top-view person ReID. To address this problem, we propose a novel approach that learns angle-related features through angular metric learning. To achieve this goal, we incorporate angular information by encoding the input capture angle in an embedding. In addition, we propose a novel angular metric loss to monitor the effect of viewpoint variations. To facilitate research in top-view person ReID, we contribute a top-view person dataset called TVReID, which contains 9168 top-view images from 148 identities. Extensive experimental results demonstrate the effectiveness of the proposed method, which achieves state-of-the-art performance on TVReID.