Integration of Multi-Head Self-Attention and Convolution for Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification is crucial for intelligent video analytics and downstream tasks.
• Existing models often overlook computational complexity, hindering practical deployment.
• Self-attention mechanisms show promise for enhancing feature representation in re-identification.

Purpose of the Study:

• To develop an improved multi-scale feature learning structure (DM-OSNet) for person re-identification.
• To enhance both accuracy and computational efficiency compared to existing models like OSNet.
• To address the practical deployment challenges of person re-identification systems.

Main Methods:

• Replaced the 9x9 convolutional stream in OSNet with double-layer multi-head self-attention to reduce computational complexity.
• Implemented a multi-scale feature learning structure (DM-OSNet) for discriminative feature extraction.
• Utilized unsupervised pre-training with SpCL on the LUPerson dataset for improved performance.

Main Results:

• DM-OSNet demonstrated superior performance over the original OSNet.
• Achieved significant reductions in computational complexity from O((H×W)2) to O(H×W×G2).
• Attained mAP scores of 87.36% (Market1501), 78.26% (DukeMTMC-reID), 72.96% (CUHK03), and 57.13% (MSMT17).

Conclusions:

• DM-OSNet offers a computationally efficient and accurate solution for person re-identification.
• The integration of multi-head self-attention effectively enhances feature learning.
• The proposed model shows strong potential for real-world intelligent video analytics applications.