ReMamba: a hybrid CNN-Mamba aggregation network for visible-infrared person re-identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible-Infrared Person Re-identification (VI-ReID) faces challenges due to intra-class variations and cross-modality differences.
• Current Convolutional Neural Network (CNN) and Vision Transformer (ViT) methods have limitations in capturing global features and managing computational complexity.

Purpose of the Study:

• To propose a hybrid network, ReMamba, for improved Visible-Infrared Person Re-identification.
• To effectively extract discriminative modality-shared features by integrating local and global information.

Main Methods:

• Utilized a CNN backbone for multi-level feature extraction.
• Introduced the Visual State Space (VSS) model to integrate local features and enhance global feature clarity.
• Designed an adaptive feature aggregation module with auxiliary loss for optimal feature fusion.
• Implemented a modal consistency identity constraint loss to reduce cross-modality differences.

Main Results:

• ReMamba demonstrated superior performance compared to state-of-the-art methods on SYSU-MM01, RegDB, and LLCM datasets.
• The hybrid approach effectively addressed limitations of existing CNN and ViT models in VI-ReID.

Conclusions:

• The proposed ReMamba framework offers a robust solution for Visible-Infrared Person Re-identification.
• Effective integration of local and global features, alongside modal consistency, is crucial for advancing VI-ReID performance.