Dynamic Weighting Network for Person Re-Identification.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Hybrid Convolution-Transformer architectures offer combined local and global feature extraction but risk losing fine-grained convolutional features when Transformers are directly embedded.
• Existing hybrid models face challenges in re-identification tasks due to the potential loss of crucial local details.
• Pure Transformer models can be computationally expensive, limiting their practical application.

Purpose of the Study:

• To propose a novel feature fusion gate unit for hybrid Convolution-Transformer networks.
• To develop a dynamic weighting network (DWNet) that effectively fuses convolutional and self-attentive features.
• To improve the performance of image re-identification models by dynamically balancing local and global feature representations.

Main Methods:

• Introduction of a feature fusion gate unit that dynamically adjusts the ratio of local and global features based on input.
• Integration of the feature fusion gate unit into different network layers and residual blocks.
• Development of the dynamic weighting network (DWNet) with ResNet (DWNet-R) and OSNet (DWNet-O) backbones.

Main Results:

• DWNet significantly enhances re-identification performance compared to baseline models.
• The proposed model maintains reasonable computational costs and parameter counts.
• DWNet-R achieved mAP scores of 87.53% (Market1501), 79.18% (DukeMTMC-reID), and 50.03% (MSMT17).
• DWNet-O achieved mAP scores of 86.83% (Market1501), 78.68% (DukeMTMC-reID), and 55.66% (MSMT17).

Conclusions:

• The feature fusion gate unit effectively addresses the limitations of directly embedding Transformers in convolutional networks.
• DWNet offers a simple, portable, and effective solution for improving image re-identification accuracy.
• The dynamic fusion of features provides superior performance across multiple benchmark datasets.