Spatiotemporal Aggregation of Appearance and Motion Features for Video-based Person Re-Identification

Video-based person re-identification (Re-ID) aims to associate similar person across multiple and non-overlapping camera views. However, most existing methods primarily focus on extracting appearance feature and often ignore the motion feature. This paper proposes spatiotemporal aggregation of appearance and motion features (STAAM) for video-based person Re-ID. More specifically, to extract discriminative appearance feature, a joint spatiotemporal feature aggregation module is proposed, which integrates spatiotemporal features at different scales to obtain appearance representation. To extract effective motion feature, a motion feature extraction module (MFEM) is presented, which extracts motion features from consecutive frames based on feature similarity. The spatiotemporal aggregation enables STAAM for extracting effective and discriminative representations for video clips, which is of benefit to recognition performance in complex scenes. Experimental results show that STAAM outperforms most existing methods on two widely used benchmark datasets and the spatiotemporal aggregation is effective for video-based person re-identification.