Bidirectional Feature Enhancement for Video-Based Occluded Cloth-Changing Person Re-Identification

Abstract

Video-based Person Re-Identification (Re-ID) has seen significant progress with deep learning techniques aimed at extracting robust spatial features from frame sequences. However, existing methods suffer from key limitations, particularly in dealing with occlusions and clothing changes, which are prevalent in real-world scenarios. In this work, we address the task termed Video-based Occluded Cloth-Changing Re-ID (VOCCRe-ID), which aims to address the compounded challenges of occlusions and clothing changes in video sequences. We propose Bidirectional Feature Enhancement (BiFE), a method that uses a novel occlusion synthesis strategy, synthesizing occlusions only on middle consecutive frames and allowing their positions to shift dynamically, thereby mimicking more realistic occlusion scenarios. Our method utilizes non-occluded keyframes at both ends of the sequence for bidirectional feature propagation, ensuring temporal consistency and filling occlusion gaps. We also introduce an attention mechanism to correct noise and ambiguity in the propagated features, which are then used to enhance both appearance and shape features. To provide a more practical benchmark, we rebuild the Occluded-VCCR dataset using the proposed occlusion synthesis strategy, making it better approximate real-world occlusion patterns, despite being synthetically generated. Experimental results demonstrate significant improvements in VOCCRe-ID performance, showing that our approach effectively handles occlusions and clothing changes while ensuring robust feature representation.