Set Residual Network for Silhouette-Based Gait Recognition
Saihui Hou, Xu Liu, Chunshui Cao, Yongzhen Huang
Abstract
Recently gait recognition receives increasing attention since it can be conducted at a long distance without the cooperation of subjects and suit for the cases of changing clothes. A key challenge is to learn gait features from the silhouettes that are invariant to the external factors such as clothing, carrying conditions and camera viewpoints. In this work, we propose a Set Residual Network for gait recognition which tries to learn more discriminative features from the silhouettes. Specifically, the silhouettes of each gait sequence are regarded as an unordered set and we propose a Set Residual Block to extract the silhouette-level and set-level features in a parallel way. Particularly, a residual connection is adopted to connect the two-level features inside each block, which enables more silhouette-set interaction and effectively coordinates the silhouette-level and set-level information for set-based feature learning from the silhouettes. Besides, we propose an efficient strategy to exploit the features from the shallow layers to learn more robust part representations for gait recognition, where the upsampling or lateral connections are unnecessary and only marginal memory cost is required. The experiments on CASIA-B and OUMVLP show that our approach can bring consistent improvement over the baselines for all walking conditions.