Exploring global-local feature fusion for pedestrian re-identification

Abstract

Pedestrian re-identification (ReID) is a challenging problem in computer vision, crucial for surveillance and security applications. Despite significant advancements, existing methods like Identity Discriminative Embedding (IDE) and Part-based Convolutional Baseline (PCB) have limitations. IDE captures global features but lacks detailed local information, while PCB focuses on local details without considering the global context. To address these issues, this paper attempts a fused model combining the strengths of both approaches. The fused model integrates global features from IDE and part-based features from PCB, creating a comprehensive representation that captures both holistic and localized details. Experimental results on Market-1501 and DukeMTMC datasets demonstrate the fused model's improved performance, significantly outperforming IDE and achieving comparable results to PCB. The fused model offers a balanced solution, apparently enhancing IDE performance and maintaining low computational consumption.