Learning from multi-view fragments: An adaptive consistency distillation framework for occluded person re-identification

Abstract

Occluded person re-identification (ReID) faces an “information incompleteness paradox”: a single occluded view misses discriminative cues and yields ambiguous representations, while exploiting multi-view observations typically requires multi-branch inference with high computational cost. To address this dilemma, we propose Multi-View Consistency Distillation (MVCD), a framework formulated under the Learning Using Privileged Information (LUPI) paradigm. Specifically, we construct a training-only teacher that has privileged access to multi-view fragments and identity annotations (available only during training), and transfer this privileged knowledge to a standard single-view student through consistency distillation. The teacher contains three training-only mechanisms: (1) Saliency-Guided Feature Purification (SGFP) to suppress occlusion-induced noise with label guidance; (2) Cross-View Patch Alignment (CVPA) to exploit patch correspondences for spatial rectification across views; and (3) Reliability-Guided Aggregation (RGA) to produce a low-variance, reliable supervision target. Crucially, all auxiliary components are discarded after training, enabling the student to recover more complete representations from occluded inputs with zero extra inference-time cost. Extensive experiments on five benchmarks show consistent improvements over strong baselines. On Occluded-DukeMTMC, MVCD achieves 70.4% Rank-1 accuracy and runs at 25 ms per image, outperforming prior state-of-the-art methods.