Cross-view analysis by multi-feature fusion for person re-identification

Person re-identification aims to match people across non-overlapping camera views. One of the challenges in re-identification is cross view matching, where the gallery and query data belong to different views. This problem is difficult because the person's appearance varies greatly due to significant viewpoint and poses changes. In this paper, we perform Kernel Canonical Correlation Analysis (KCCA) for each feature type to project data from different views into a coherent subspace in which the correlations from different views are maximized with this specific feature representation. A similarity score between query and gallery images is derived from each subspace. Then the scores of different features are fused with query adaptive weights to reach a final decision. This scheme is different from current KCCA approaches which only learn one single subspace for all the features and ignore the diversity of multiple features. The effectiveness of the method is evaluated on two publicly available datasets (VIPeR and PRID), yielding state-of-the-art performance with respect to recent techniques.