MHSA-Net: Multihead Self-Attention Network for Occluded Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) is crucial for security and surveillance.
• Existing models struggle with occluded images and irrelevant information.
• There is a need for more robust and accurate person Re-ID methods.

Purpose of the Study:

• To introduce a novel person re-identification model, MHSA-Net.
• To enhance the capture of key local information and prune unimportant details.
• To improve performance in both standard and occluded person Re-ID tasks.

Main Methods:

• Developed the Multi-Head Self-Attention Network (MHSA-Net).
• Incorporated a Multi-Head Self-Attention Branch (MHSAB) to capture local features.
• Implemented an Attention Competition Mechanism (ACM) to filter noise.

Main Results:

• MHSA-Net effectively prunes unimportant information.
• MHSAB adaptively captures key local person information for matching.
• ACM successfully filters attention noise and non-key information.
• Ablation studies confirmed the contribution of MHSAB and ACM.

Conclusions:

• MHSA-Net achieves competitive performance on person Re-ID benchmarks.
• The novel components significantly contribute to the model's effectiveness.
• MHSA-Net shows promise for real-world surveillance applications.