MambaReID: Exploiting Vision Mamba for Multi-Modal Object Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-modal object re-identification (ReID) leverages complementary information across different image modalities.
• Traditional Convolutional Neural Network (CNN) methods have limited receptive fields.
• Transformer-based approaches face high computational demands and lack convolutional biases.

Purpose of the Study:

• To propose a novel fusion framework, MambaReID, that integrates the strengths of CNNs and Transformers for improved multi-modal object ReID.
• To overcome the limitations of existing methods by incorporating the effective VMamba architecture.

Main Methods:

• Developed MambaReID, a fusion framework comprising Three-Stage VMamba (TSV), Dense Mamba (DM), and Consistent VMamba Fusion (CVF).
• TSV captures global context and local details efficiently with low computational complexity.
• DM enhances feature discriminability through dense integration of inter-modality information.
• CVF provides granular modal aggregation for improved feature robustness.

Main Results:

• MambaReID achieves superior performance in multi-modal object ReID tasks.
• The framework demonstrates reduced parameters and lower computational costs compared to existing methods.
• Effectiveness validated through extensive experiments on three multi-modal object ReID benchmarks.

Conclusions:

• The proposed MambaReID framework effectively addresses limitations of CNN and Transformer-based methods in multi-modal object ReID.
• MambaReID offers a computationally efficient and high-performing solution for cross-modal object identification.