Abstract
Occluded person re-identification (re-ID) remains a significant challenge due to the presence of obstacles that obscure discriminative features. Traditional methods rely either on noisy external semantic partitions or difficult-to-train attention mechanisms. This paper introduces Semi-attention Partition (SAP), a framework that leverages an "imperfect" semantic teacher to guide a Vision Transformer-based student. By balancing cooperation with "reasonable conflict," SAP enables the student to surpass the teacher's noisy guidance, achieving state-of-the-art performance on major benchmarks including Occluded-DukeMTMC.
Introduction
Person re-identification (re-ID) is the task of identifying the same individual across different camera views. In real-world scenarios, pedestrians are often partially hidden by obstacles like cars, trees, or other people—a condition known as occluded re-ID.
Current solutions generally fall into two categories: external semantic partition and attention-based partition. External semantic methods use off-the-shelf models (like human parsers) to define body parts, but these often fail on low-resolution or heavily occluded images, providing "noisy" labels. Conversely, attention-based methods use self-attention mechanisms (like Transformers) to automatically find important parts but are difficult to train from scratch on limited re-ID datasets. SAP bridges this gap by using noisy semantic guidance to help an attention-based student reach its full potential without being constrained by the teacher's errors.
Methodology
The SAP framework is built on a Vision Transformer (ViT) backbone and introduces "semi-attention"—a hybrid of standard self-attention and distilled attention guided by the teacher. The training process occurs in two distinct stages:
- Stage 1: Global Feature Learning: The model learns relationships between all image patches (patch-to-patch consistency). It uses the teacher's labels to encourage patches belonging to the same body part to have similar attention scores.
- Stage 2: Part Feature Learning: The model uses "part proxies" (learnable vectors representing body parts) to "absorb" information from relevant image patches (proxy-to-patch consistency).
To ensure the student surpasses the noisy teacher, the authors introduce two "conflict-promoting" designs: Mutual Absorbing Partition Refinement (MAPR), which allows patches and part proxies to refine each other iteratively, and Supervision Signal Dropout (SSD), which randomly drops teacher signals to force the student to rely on its own learned internal similarities.
Key Insights and Results
A central contribution of this work is the "Weak Teacher" Paradox. Traditionally, knowledge distillation involves a strong teacher and a weak student. This paper demonstrates that a "stronger" learner (a Transformer) can use a "weaker" teacher (noisy semantic labels) as a starting point and eventually surpass it. By allowing "reasonable conflict" through SSD and MAPR, the student uses the teacher for structural guidance but uses its own attention mechanism to "clean up" the details.
SAP achieved new state-of-the-art (SOTA) results across major benchmarks:
- Occluded-DukeMTMC: Reached 62.2% mAP and 70.0% Rank-1 accuracy, significantly outperforming previous methods.
- Occluded-REID: Achieved 76.8% mAP and 83.0% Rank-1.
- Market-1501: Hit 90.5% mAP, proving that the method improves general re-ID performance by learning better part alignments even in non-occluded scenarios.
Conclusion
SAP demonstrates that imperfect semantic guidance can be a powerful tool for training attention-based models in data-constrained environments. By fostering a balance of cooperation and conflict, the student becomes an independently competent feature extractor that no longer requires the teacher at test time. Future work may explore applying this semi-attention paradigm to other tasks where noisy structural priors are available, such as object detection or medical imaging.
Summary updated