A Multimodule Collaborative Framework for Unsupervised Visible-Infrared Person Re-Identification with Channel...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised visible-infrared person re-identification (USL-VI-ReID) is crucial for intelligent surveillance but faces challenges from modality gaps and limited feature granularity.
• Channel augmentation (CA) is typically used for data augmentation, with its potential as an independent input modality unexplored.

Purpose of the Study:

• To present a multimodule collaborative USL-VI-ReID framework that treats CA as a separate input modality.
• To address modality gaps and enhance feature representations in USL-VI-ReID.

Main Methods:

• A four-module framework combining Person-ReID Adaptive Convolutional Block Attention (PA-CBAM), Varied Regional Alignment (VRA) with Multimodal Assisted Adversarial Learning (MAAL), Varied Regional Neighbor Learning (VRNL), and Uniform Merging (UM).
• PA-CBAM extracts features using a two-level attention mechanism.
• VRA performs cross-modal regional alignment and MAAL reinforces region-level correspondence.
• VRNL stabilizes pseudo-labels and captures local structure via multi-region association.
• UM merges clusters using alternating contrastive learning for improved consistency.

Main Results:

• Achieved Rank-1 accuracy of 93.34%, mAP of 87.55%, and mINP of 76.08% on the RegDB dataset (visible-to-infrared).
• Demonstrated effective reduction of modal discrepancies and increased feature discriminability.
• Outperformed most existing unsupervised baselines and several supervised approaches.

Conclusions:

• The proposed framework effectively enhances USL-VI-ReID performance by treating CA as an independent modality.
• The method significantly advances the practical applicability of USL-VI-ReID in intelligent surveillance and public safety.