Context-aided unicity matching for person re-identification

Most existing person re -identification methods compute the matching relations between person images based on the similarity ranking. It lacks the global viewpoint and context consideration, inevitably leading to ambiguous matching results and sub-optimal performance. Based on a natural assumption that images belonging to the same identity should not match with images belonging to different identities across views, called the unicity of person matching on the identity level, we propose an end -to -end person unicity matching architecture to learn and refine the person matching relations. We first employ the image samples' context in feature space to generate the initial soft matchings using graph neural networks, and then utilize the samples' global context to refine the soft matchings and reach the matching unicity by bipartite graph matching. Considering real -world applications, we further develop a fast algorithm without losing performance. Experimental results on five public benchmarks show the superiority of the proposed method.