High order Interaction and Wavelet Convolution Network for visible infrared person reidentification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible-infrared person re-identification (VI-ReID) faces challenges from cross-modal discrepancies and low image quality.
• Current deep learning methods often lose crucial primitive features during high-level abstraction.

Purpose of the Study:

• To propose the High-order Interaction and Wavelet Convolution Network (HIW-Net) to address information loss in VI-ReID.
• To enhance feature extraction by integrating primitive features and employing wavelet convolution.

Main Methods:

• Developed HIW-Net, integrating primitive features at multiple interaction stages.
• Utilized wavelet convolution for diverse feature mining and improved feature extraction.
• Created the RegDB_shape dataset using the Segment Anything Model (SAM) for training augmentation.

Main Results:

• HIW-Net demonstrated superior performance compared to state-of-the-art methods on SYSU-MM01 and RegDB datasets.
• The proposed method effectively compensates for information loss in high-order representations.
• Wavelet convolution contributed to more comprehensive feature extraction.

Conclusions:

• HIW-Net offers a significant advancement in visible-infrared person re-identification.
• The integration of primitive features and wavelet convolution proves effective for cross-modal discrepancies.
• The RegDB_shape dataset aids in improving VI-ReID model training.