Structure-Aware Positional Transformer for Visible-Infrared Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible-infrared person re-identification (VI-ReID) addresses cross-modality pedestrian matching challenges.
• Existing methods struggle with pose variation, occlusion, and inter-modality differences, often relying on global or uniform part features sensitive to misalignments.

Purpose of the Study:

• To develop a novel network, the structure-aware positional transformer (SPOT), for learning semantic-aware sharable modality features in VI-ReID.
• To improve robustness against misalignments and reduce cross-modality differences.

Main Methods:

• The proposed SPOT network integrates an attended structure representation (ASR) and transformer-based part interaction (TPI).
• ASR models modality-invariant structure features and selects discriminative regions guided by structure.
• TPI utilizes transformers to mine part-level appearance and position relations for learning discriminative features.

Main Results:

• SPOT effectively utilizes contextual and structural information, reducing cross-modality differences.
• The method demonstrates superior performance over state-of-the-art approaches on two cross-modal datasets.
• Significant improvements were observed, including an 8.43% increase in Rank-1 and 6.80% in mAP on the SYSU-MM01 dataset.

Conclusions:

• The SPOT network offers an effective solution for VI-ReID by incorporating structural and positional information.
• This approach enhances robustness against common challenges like pose variation and occlusion.
• SPOT represents a significant advancement in cross-modality person re-identification accuracy and reliability.