Graph similarity rectification for person search

Abstract In person search task, it is hard to retrieve the query persons undergoing large visual changes. To tackle this problem, we propose to exploit the context information to rectify the original individual similarity for better retrieval. To this end, we propose to model a query frame and a gallery frame as a graph pair, and then design the Siamese Residual Graph Convolutional Networks (SR-GCN) to aggregate context information to generate graph similarity as a complement of the original similarity. To model the relationships between context persons, we define the joint similarity adjacency matrix which assigns the proposed joint similarity as the edge weight to measure the contributions a context person makes to the aggregation. Therefore, the context node with a higher possibility to be a co-traveler of the target node makes more contributions to the matching of the target node. To further enhance the discriminative power of individual features, we also design a Random Proxy Center loss which explicitly constrains the intra-class variations to be smaller than the inter-class variations in the feature space and could make use of unlabeled samples. Experimental results on two public datasets show that our approach performs favorably against the state-of-the-art methods.