A local-global transformer-based model for person re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (ReID) is crucial for recognizing individuals across multiple camera views.
• Both Transformer and Convolutional Neural Network (CNN) based methods achieve competitive performance in ReID.
• Transformer models often neglect local features by processing sequences holistically.

Purpose of the Study:

• To develop an improved Transformer-based person ReID model.
• To effectively integrate local and global features for enhanced recognition accuracy.
• To address the limitation of Transformer models overlooking local image details.

Main Methods:

• Introduced a novel Transformer-based person ReID model.
• Incorporated a Local Attention Module to capture fine-grained local features.
• Integrated relative position encoding within the Local Attention Module to better understand structural information.

Main Results:

• The proposed model demonstrated improved performance on person ReID tasks.
• Achieved a Rank-1 accuracy improvement of 0.7% on the Market-1501 dataset.
• Achieved a Rank-1 accuracy improvement of 0.9% on the DukeMTMC-reID dataset.

Conclusions:

• The integration of local and global features enhances Transformer-based person ReID.
• The Local Attention Module with relative position encoding effectively captures crucial image structure.
• The model shows significant improvements on standard person ReID benchmarks.