TE-TransReID: Towards Efficient Person Re-Identification via Local Feature Embedding and Lightweight Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) matches individuals across non-overlapping cameras.
• Transformer models excel in Re-ID but suffer from high computational costs and limited local feature capture.
• Existing methods struggle to balance accuracy and efficiency in person Re-ID.

Purpose of the Study:

• To propose a novel Toward Efficient Transformer-based Person Re-identification (TE-TransReID) framework.
• To address the computational expense and local feature limitations of current Transformer-based Re-ID models.
• To achieve a balance between high recognition accuracy and lightweight network design.

Main Methods:

• Utilizes the initial layers of a pretrained Vision Transformer (ViT) for global features.
• Extracts local features using a pretrained Convolutional Neural Network (CNN).
• Employs a dual efficient feature-fusion strategy (ETFFM and EPFFM) to integrate global and local features.

Main Results:

• Achieved rank-1 accuracy of 94.8% on Market1501, 88.3% on DukeMTMC, and 85.7% on MSMT17.
• Maintains comparable recognition accuracy to existing hybrid models with significantly fewer parameters (27.5 M).
• Demonstrates a drastic reduction in model parameters while preserving high performance.

Conclusions:

• The TE-TransReID framework establishes an optimal equilibrium between recognition accuracy and computational efficiency.
• This approach offers a practical solution for efficient person re-identification tasks.
• The proposed method effectively integrates global and local features for improved performance.