Multi-granularity dynamic hierarchical graphs for video-based person re-identification

Video-based person re-identification (Re-ID) aims to identify video sequences of specific pedestrians captured in a distributed camera system. The key to tackling this challenging task is to model rich temporal features in video sequences. However, existing methods usually capture only fixed-length global temporal clues, making it difficult to eliminate the effect of occlusion and accurately model the motion information of pedestrians. To address this limitation, we propose a novel graph-based framework, namely dynamic hierarchical graph network(DHGN), to model temporal features by adaptively capturing multi-granularity temporal clues. Specifically, we adaptively segment the video features horizontally into several regions and construct graphs for the features at the same level. The features of each frame at the same level are considered as graph nodes, which are then adaptively connected based on feature similarity. By dynamically aggregating features from neighboring nodes in the same graph, DHGN is able to adaptively capture temporal cues from different body parts. Furthermore, in order to obtain more robust matching results, we propose a similarity weighted inference module (SWIM), which utilizes gallery-gallery similarity to modify the query-gallery similarity matrix. Extensive experiments on four benchmarks clearly demonstrate the effectiveness of the proposed method.