Attribute-driven Temporal Attention for VIdeo-based Pedestrian Analysis

ON2019-05-15文章来源 信息科学与技术学院CATEGORY全部

Speaker: Prof. Annan Li

Time: 15:30-16:30, May 15

Location: SIST 1A 200

Host: Prof. Shenghua Gao


Li, Annan is currently an associate professor with School of Computer Science and Engineering, Beijing Univesity of Aeronautics and Astronautics. Prior to that, he spent five years at Singapore, worked as a scientist with Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR) and as a postdoctoral research fellow at National University of Singapore respectively. He currently works with the Intelligent Recognition and Image Processing Laboratory. His major research interests include computer vision, pattern recognition, and biometrics.


Video-based pedestrian analysis task such as person re-identification (Re-ID) and attribute recognition are important issues in computer vision, where one of the key challenges is reducing the negative influence introduced by unstable human detector/tracker, background clutter and surrounding people. The mitigation of noisy frame can be achieved by frame weighting which is also known as the temporal attention in deep neural networks. Different from conventional approaches, we learn the attention from middle-level attributes instead of identity label. The proposed multi-task framework is referred as GraftNet, since the person Re-ID is implemented by grafting attribute-based temporal attentions to an identity recognition network. Both identity-relevant and irrelevant attribute are investigated. Optimizing the former and mitigating the latter can both improves the temporal attention. Extensive experiments on MARS and DukeMTMC-VID show that the proposed model is very effective and outperforms the state-of-the-arts.