Focus on the Visible Regions: Semantic-Guided Alignment Model for Occluded Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian retrieval is hindered by occlusion, where obstacles obscure parts of individuals.
• Existing re-identification (Re-ID) methods often fail to adequately address occlusion noise.
• This noise degrades the accuracy and reliability of retrieval systems.

Purpose of the Study:

• To develop a novel semantic-guided alignment model for robust pedestrian re-identification (Re-ID) under occlusion.
• To effectively filter out noise introduced by occluded regions in images.
• To enhance the performance of holistic Re-ID systems.

Main Methods:

• Utilized a human semantic parsing network to generate probability maps identifying occluded image regions.
• Implemented automatic image cropping based on probability maps to preserve visible human parts.
• Fused probability maps with global image features to guide local feature extraction on visible regions.
• Introduced a measurement strategy focusing on the distance between visible human areas (public areas) for matching.

Main Results:

• The proposed model successfully separates useful information from occlusion noise.
• Experimental results demonstrate superior performance compared to previous occluded Re-ID methods.
• The method achieved top performance in holistic Re-ID benchmarks.

Conclusions:

• The semantic-guided alignment model offers a significant advancement in handling occluded pedestrians.
• This approach effectively mitigates the impact of occlusion noise on Re-ID accuracy.
• The method provides a robust solution for real-world pedestrian retrieval challenges.