Domain adaptive person re-identification with noise optimization and dynamic weighting

Abstract

Domain adaptive person re-identification (Re-ID) faces challenges due to inherent noise from limited domain transferability and the uncertainty in pseudo-label generation. To address this, we propose NODW (Noise Optimization and Dynamic Weighting), a comprehensive domain adaptive person Re-ID framework that systematically tackles these issues through quantitative noise assessment and dynamic optimization. Our method proposes: (1) an enhanced ResNet50-pro backbone specifically designed for cross-domain feature extraction, (2) a silhouette coefficient-based module for pseudo-label quality assessment with dynamic weighting, (3) a Maximum Mean Discrepancy (MMD)-based module for minimizing domain transferability limitations, and (4) a robust consistency supervision mechanism to ensure stable feature learning. Extensive experiments demonstrate state-of-the-art performance across multiple domain transfer tasks, achieving mAP scores of 73.8% (Market to Duke), 84.7% (Duke to Market), 34.2% (Market to MSMT), and 35.6% (Duke to MSMT). These results represent significant improvements over existing methods, particularly in challenging scenarios with large domain gaps, validating the effectiveness of our noise-aware adaptation strategy.