Re-ranking Person Re-identification with Local Discriminative Information

Most existing metric learning based person reidentification methods try to learn a global distance metric to measure the similarity between person images. But owing to the large intra-class variations, pedestrian data follows very irregular distribution in the feature space. The global metric model can hardly exploit the discriminative information from local distribution. Thus, due to the higher similarity of distribution, local information should be elaborately mined and exploited to improve the matching accuracy, especially for some hard positive images. In this paper, we propose to combine the global metric and local information to resolve failure matching cases. Detailly, for a testing pair, positive pairs in the training set whose feature differences are similar with given testing pair under global metric are firstly searched. If most of these positive pairs are located in the local range of the testing pair, the global metric is thus believed to reflect the similarity relationship in this local area. According to the degree of local discriminative information being represented in global metric, testing pair is derived based on the global metric as well as the given pair's local information. Finally, all gallery images are re-ranked according to the combined similarity scores. Experimental results on VIPeR, PRID450S and Market-1501 datasets clearly demonstrate the effectiveness of the proposed method.