Attention-Guided Siamese Network for Clothes-Changing Person Re-identification.

Person re-identification (re-id) has achieved significant progresses in recent years. However, the existing methods generally assume that the clothes of pedestrians remain unchanged throughout the surveillance periods, which is contradict to realistic environment where pedestrians may change their clothes. Current re-id techniques may encounter a dramatic performance degradation when the pedestrians change the clothes. In this paper, we propose a novel attention-guided siamese network (AGS-Net) to solve the cross-clothes re-id challenge. The AGS-Net integrates the visual and contour information together by developing a dual-branch structure, among which one extracts powerful features from raw inputs while the other learns robust features from the sketch image. Moreover, we exploit the attention modules to emphasize reliable identity-related features considering changing clothes and avoid generating features sensitive to clothes. Specifically, we propose a clothes-change invariant constraint to learn clothes-invariant features. Experimental results verify the effectiveness of our approach.