Multi-level Semi-Coupled Dictionary Learning for Enhanced Person Re-Identification

Person Re-ID focuses on developing a robust feature matching framework that effectively associates images of the same individual captured from different camera per-spectives. In this work, we introduce a multi-level feature matching framework for person re-identification that leverages both local and global visual cues. The framework stratifies the re-identification task into two concurrent pipelines. Within these pipelines, this paper presents a pair of novel semi-coupled dictionary learning (SCDL) algorithms to develop distinct yet complementary feature representations: (1) The Local SCDL (L-SCDL) algorithm focuses on region-level descriptors, mapping them to a mutual feature space that diminishes the influence of camera-specific variations. (2) The Global SCDL (G-SCDL) algorithm, on the other hand, is designed to harmonize global appearance attributes, thereby enhancing the robustness of image-level descriptor matching. These two complementary components synergistically exploit both global and local visual cues, thereby enhancing the discriminative ability of our approach. Extensive experiments conducted on three challenging datasets demonstrate the superior performance of the proposed person re-identification framework compared to several mainstream approaches.