Information Bottleneck Enhanced Video-based Person Re-identification

Yaqiao Liao, Xu Wang, Guowen Kuang, Zhigang Li, Tingting Li

Abstract: Person re-identification has great potential in applications of public security. However, non-discriminative information like background noise still hinders further improvement of current person re-identification algorithms. This paper proposes an information bottleneck-enhanced video-based person re-identification algorithm concentrating on discriminative information. Specifically, the spatial-temporal information purification (STIP) module is designed to filter out irrelevant information and the variance of information is estimated to weigh the importance of each feature dimension. Our method is evaluated on MARS and results show that it pays more attention to discriminative information and suppresses the irrelevant feature.