AAGCN: Adjacency-aware Graph Convolutional Network for person re-identification

摘要

Person re-identification (ReID) is an important topic of computer vision. Existing works in this field focus primarily on learning a feature extractor that maps the pedestrian images into a feature space, in which feature vectors corresponding to the same identity are close to each other. In this paper, we propose the adjacency-aware Graph Convolutional Network (AAGCN) to smooth the intra-class features and thus reduce the intra-class variance. Specifically, our AAGCN takes the features learned by a backbone as the input nodes; it first establishes the connections or adjacency relations for the intra-class features, then the adjacent nodes (i.e., the intra-class features) would be smoothed thanks to the property of low-pass filtering of Graph Convolutional Network (GCN). In this paper, we propose two methods, i.e., the Mahalanobis Neighborhood Adjacency (MNA) and Non-Linear Mapping (NLM), to learn the adjacency relations for the intra-class features. The MNA defines the adjacency weight between two nodes as the negative exponent of the Mahalanobis distance between their corresponding features, therefore it aims to learn a small Mahalanobis distance between the intra-class features and a large Mahalanobis distance between the inter-class ones. The NLM enables the non-linear mapping from the features of the nodes to their corresponding adjacency weights. The experimental results on both visible ReID and visual–infrared ReID verify the effectiveness of our method, for instance, our model achieves 95.7% rank-1 and 93.1% mAP on Market1501, as well as 58.6% rank-1 and 60.0% mAP on SYSU.