GiT: Graph Interactive Transformer for Vehicle Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Transformers are increasingly used in computer vision for global feature extraction from image patches.
• Pure transformers lack discriminative local features crucial for vehicle re-identification.
• Vehicle re-identification demands both robust global and discriminative local features.

Purpose of the Study:

• To propose a novel Graph Interactive Transformer (GiT) model for vehicle re-identification.
• To enhance feature extraction by integrating local and global feature learning.
• To improve the accuracy and robustness of vehicle re-identification systems.

Main Methods:

• A stacked architecture of GiT blocks is employed, combining graph-based local feature extraction and transformer-based global feature extraction.
• An interactive mechanism between graph and transformer components is implemented at each level.
• A novel local correction graph module is introduced to learn discriminative local features within patches by analyzing node relationships.

Main Results:

• The proposed GiT method demonstrates superior performance compared to state-of-the-art approaches.
• Experiments conducted on three large-scale vehicle re-identification datasets validate the effectiveness of the GiT model.
• The integration of local and global features through graph-transformer interaction yields significant improvements.

Conclusions:

• The Graph Interactive Transformer (GiT) effectively addresses the limitations of pure transformers in vehicle re-identification.
• GiT achieves state-of-the-art results by synergistically leveraging local and global features.
• The proposed local correction graph module enhances the learning of discriminative local features.