Transformer-based vehicle re-identification with view information.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• The industrial Internet of Things and smart cities necessitate advanced vehicle re-identification (ReID) systems.
• Key challenges in ReID include intra-identity variations (different views of the same vehicle) and inter-identity similarities (similar views of different vehicles).

Purpose of the Study:

• To propose a novel Transformer-based Vehicle ReID model with View Information (TVRVI) to address the challenges of intra-identity differences and inter-identity similarities.
• To improve the accuracy and robustness of vehicle re-identification in large-scale datasets.

Main Methods:

• Annotated vehicle images into five view parts and trained a parsing network to assign view labels.
• Designed a dual-branch Transformer-based parsing network to separately extract features from different views, minimizing entanglement.
• Utilized view information to enable the local branch to learn fine-grained view representations and the global branch to learn overall image features.

Main Results:

• The TVRVI model effectively leverages common-view features while mitigating interference from uncommon-view features.
• Demonstrated significant reductions in intra-identity differences and inter-identity similarities between vehicle images.
• Achieved state-of-the-art performance on four public vehicle datasets, validated by ablation studies.

Conclusions:

• The proposed TVRVI model, incorporating view information, significantly enhances vehicle re-identification accuracy.
• The dual-branch Transformer architecture effectively processes multi-view features for improved ReID performance.
• View information is crucial for overcoming common challenges in vehicle re-identification tasks.