LAF: Enhancing person re-identification via Latent-Assisted Feature Fusion

Abstract

Person re-identification (Re-ID) in real-world scenarios is challenged by occlusions, viewpoint variations, and individuals with similar attributes. Existing methods predominantly rely on salient regions, yet such regions often become unreliable under occlusion or in crowded environments, leading to ambiguous feature representations. To address this limitation, we propose a novel Latent-Assisted Fusion (LAF) framework that systematically mines discriminative cues from non-salient areas, which are critical for distinguishing challenging samples. Our approach introduces three key innovations: Lock-Drop, Outlook-Attention, and ML-Fusion. Lock-Drop selectively erases prominent regions based on primary features, encouraging the model to learn from less obvious areas. Outlook-Attention refines the latent information, while ML-Fusion integrates these enriched features with the primary ones, significantly boosting the robustness and diversity of the learned features. Extensive experiments on five large-scale person re-identification benchmarks demonstrate that LAF consistently improves the performance of existing algorithms. Compared to state-of-the-art methods, LAF achieves superior results, including an mAP of 89.6% and Rank-1 accuracy of 95.9% on the Market1501 dataset, and an mAP of 63.3% with Rank-1 accuracy of 84.1% on the MSMT17 dataset. These results highlight the effectiveness of our proposed module in leveraging latent information from non-salient regions, leading to substantial performance improvements, particularly in challenging scenarios involving occlusions and complex backgrounds. Code is available at https://github.com/meanlang/LAF.