Social Interaction Modeling for Group Re-identification

Xiaoming Guo, Ping Hu, Jingke Meng, Jian Zhao, Junhang Wu, Kai Zhao

Keywords: Group Re-identification; Member Differentiation; Social Interaction; Attention Mechanism; Layout Variation

TL;DR: Greid

Abstract: Group Re-identification (G-ReID) focuses on associating group images that contain the same members across different camera views. The key challenge is that identity and position differentiation in group topology structure changes are difficult to capture. Drawing on principles from social psychology, we observe that the core members are more likely to remain in the group under different camera views with smaller position changes, while peripheral members are more likely to have significant position changes or even fade out of the group. To this end, we propose a novel social interaction modeling (SIM) method, which treats each group as a social interaction field to explore more authentic and robust group features through dealing with the member differentiation: identity and position differentiation. Our method constructs the social interaction calculation module (SICM) to capture the member differentiation in the fields, and implements identity differentiation and position differentiation by the social prior attention mechanism (SPAM) and social layout variation module (SLVM), respectively. Extensive experiments on three available datasets show that the proposed method SIM is effective, and outperforms all previous state-of-the-art methods, surpassing the baseline on Rank1/mAP by up to 8.6\%/9.6\% on DukeGroup, 3.7\%/2.7\% on RoadGroup and 2.5\%/2.9\% on CSG. The code will be available on Github.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18002