Salient Region-Based Least-Squares Log-Density Gradient Clustering for Image-To-Video Person Re-Identification

Abstract

Because of its distinction and reliability, the person-salient region has been applied to pedestrian re-identification (re-id) across disjoint camera views. Despite the great progress achieved, a few works have studied the more challenging image-to-video person re-id problem, in which the gallery consists of videos and the same pedestrian appears in a continuous video sequence. The intrinsic high redundancy of such video sequences makes it more difficult to obtain high performance for pedestrian re-id. To solve this problem, in this paper, we propose a new salient region clustering approach for image-to-video person re-id. Specifically, we use the mean shift to extract the person regions and obtain the salient region by computing the saliency for each region. Then, all salient regions are clustered by least-squares log-density gradient clustering, after which the salient regions that are extracted from the same person are marked as the same group. Finally, the rank of the person re-id can be obtained by computing the distance between the probe-salient region and the gallery clustered salient regions. We evaluate the proposed approach on two public pedestrian sequence data sets (PRID 2011 and MARS), and the experimental results validate the effectiveness of the proposed approach for the image-to-video person re-id.