Abstract
Person re-identification (re-ID) based on visual appearance has been an intensively researched area in computer vision and forensic multimedia analysis. Its goal is to associate person detections under different spatial-temporal scenarios across different camera views. Existing efforts on person re-ID can generally be categorized into two approaches: conventional image retrieval and highly-crafted re-ID structures. In this paper, we formulate person re-ID, for the very first time, as an energy-based deep structured prediction problem without the need of explicitly specifying the graph topology of the re-ID structure in advance. We also integrate a structure sampling mechanism, Randomized Dropout Structure Sampling (RDSS), into structured prediction while all the existing works assume that structure samples are readily available for learning. Experiment results show that our new formulation outperforms conventional image retrieval and highly crafted re-ID structures.