Multi-Level Fusion Temporal-Spatial Co-Attention for Video-Based Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) struggle with local minima and unstable training due to insufficient data in person re-identification.
• Existing methods often require laborious collection of auxiliary data like pedestrian attributes and postures.
• Variations in frame similarity within video sequences pose challenges for accurate re-identification.

Purpose of the Study:

• To improve video-based person re-identification (reID) performance, particularly for small datasets.
• To develop a method that mitigates overfitting and reduces dataset limitations.
• To enhance the generalization ability and accuracy of reID networks.

Main Methods:

• Implemented a multi-level fusion temporal-spatial co-attention mechanism.
• Introduced the concept of knowledge evolution to enhance a backbone Residual Neural Network (ResNet).
• Utilized parallel global, local, and attention branches for feature extraction, embedding high-level features into a metric learning network.

Main Results:

• The improved network demonstrated enhanced ability to prevent overfitting on small datasets like PRID2011 and iLIDS-VID.
• Experiments on MARS and DukeMTMC-VideoReID showed improved feature extraction and generalization.
• Achieved 90.15% Rank1 and 81.91% mAP on the MARS dataset.

Conclusions:

• The proposed multi-level fusion temporal-spatial co-attention method significantly improves video-based person re-identification.
• The approach effectively addresses challenges posed by small datasets and enhances model generalization.
• The method offers a more robust and accurate solution for person re-identification tasks.