A Multi-Level Relation-Aware Transformer model for occluded person re-identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification (Re-ID) is a significant challenge due to occlusions from objects or other pedestrians.
• Existing Re-ID methods often rely on auxiliary models like pose estimation, which can be unreliable with occlusions.
• Previous approaches frequently learn features from single images, neglecting inter-sample relationships crucial for robust Re-ID.

Purpose of the Study:

• To develop a novel Multi-Level Relation-Aware Transformer (MLRAT) model for enhanced occluded person Re-ID.
• To address limitations of auxiliary models and single-image feature learning in existing Re-ID techniques.
• To improve the accuracy and robustness of person Re-ID in scenarios with significant occlusions.

Main Methods:

• Introduced the Multi-Level Relation-Aware Transformer (MLRAT) model, comprising Patch-Level Relation-Aware (PLRA) and Sample-Level Relation-Aware (SLRA) modules.
• PLRA utilizes a Graph Convolutional Network (GCN) to model structural relations between key image patches, bypassing auxiliary models.
• SLRA employs a Relation-Aware Transformer (RAT) block and self-distillation to model inter-sample relationships and transfer knowledge.

Main Results:

• The MLRAT model demonstrated significant performance improvements over existing baselines on four occluded person Re-ID datasets.
• The model maintained competitive performance on partial and holistic person Re-ID datasets.
• PLRA effectively learned local features by modeling patch relations, while SLRA captured discriminative sample-level features.

Conclusions:

• The proposed MLRAT model effectively addresses the challenges of occluded person Re-ID by leveraging multi-level relational modeling.
• The novel PLRA and SLRA modules provide a robust framework for learning discriminative features without reliance on auxiliary models.
• MLRAT offers a superior solution for occluded person Re-ID, showing strong generalization across various dataset types.