UD-Gaussian: Uncertainty-Driven Gaussian Modeling for Occluded Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Occluded person re-identification faces challenges due to obscured pedestrians.
• Existing methods using pose/semantic info suffer from cross-domain gaps and instability.
• Complex occlusion environments demand more robust feature learning.

Purpose of the Study:

• To develop a Transformer-based uncertainty-driven Gaussian model (UD-Gaussian) for occluded person re-identification.
• To enhance pedestrian feature representation and improve model discriminative ability under occlusion.
• To address the instability and false results of existing methods.

Main Methods:

• Introduced a high-frequency enhancement module using Discrete Haar Wavelet Transform and graph attention.
• Developed a probability distribution learning module with a memory bank and Gaussian distributions.
• Utilized entropy as a loss function to promote deterministic and independent probability distributions.

Main Results:

• The high-frequency enhancement module enriches detailed pedestrian image features.
• The probability distribution learning module enhances model discriminative ability.
• Experimental results demonstrate superior performance on occluded and holistic person re-identification datasets.

Conclusions:

• UD-Gaussian effectively handles challenges in occluded person re-identification.
• The proposed method offers improved stability and accuracy compared to existing approaches.
• The integration of high-frequency enhancement and uncertainty-driven learning is key to its success.