Title: DPM++: Dynamic Masked Metric Learning for Occluded Person Re-identification

Abstract:

While significant strides have been made in person re-identification, handling occlusions from physical obstacles remains a persistent challenge in practical deployments. The core difficulty stems from the disparity between partial, occluded observations and holistic identity features. Heavy occlusion not only eliminates distinctive body details but also injects noise from background clutter and blocking objects, rendering traditional global metric learning ineffective. Current approaches typically depend on auxiliary pre-trained models to identify visible regions for alignment or generate occluded data through augmentation techniques; however, they fail to offer a cohesive framework capable of learning robust, visibility-consistent matching under authentic occlusion conditions.

To address this, we introduce DPM++, a Dynamic Masked Metric Learning framework designed for occluded person re-identification. This approach acquires an input-adaptive masked metric that dynamically identifies trustworthy identity subspaces for each occluded sample. By doing so, it prioritizes matching based on visible, consistent evidence while minimizing the impact of unreliable features. Grounded in a classifier-prototype space, DPM++ employs a two-stage supervision method utilizing CLIP. In this scheme, semantic priors at the ID level are extracted from the text branch and mapped into the classifier-prototype space to facilitate dynamic masked matching.

To further enhance the masked metric, we propose a saliency-guided patch transfer strategy. This method synthesizes controllable, photorealistic occluded samples during the training phase. By leveraging real-world scene priors, the strategy ensures the model encounters realistic partial views, offering more informative supervision than standard random erasing techniques. Additionally, the framework incorporates occlusion-aware sample pairing and mask-guided optimization to boost stability and overall performance. Evaluations on benchmarks covering both occluded and holistic person re-identification demonstrate that DPM++ consistently surpasses existing state-of-the-art methods across both scenarios.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC