An Investigation of the Domain Gap in CLIP-Based Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-id) is crucial for intelligent surveillance.
• The domain gap, or performance drop on unseen data, hinders re-id systems.
• CLIP-based models offer potential for domain generalization via multimodal pre-training.

Purpose of the Study:

• To quantitatively analyze the domain gap in CLIP-based re-id systems.
• To evaluate CLIP's effectiveness in mitigating domain shift compared to image encoders.
• To assess the impact of data augmentation and extended training on re-id robustness.

Main Methods:

• Systematic performance evaluation of CLIP-based re-id models on standard benchmarks (Market-1501, DukeMTMC-reID, MSMT17, Airport).
• Measurement of mean average precision (mAP) and Rank-1 accuracy.
• Analysis of CLIP's visual-textual alignment benefits and comparison with image encoder baselines.
• Evaluation of training set expansion and random erasing augmentation.

Main Results:

• CLIP's visual-textual alignment offers advantages in person re-identification.
• Extending training data and using random erasing improved mAP by +4.3% and Rank-1 accuracy by +4.0% on average.
• The study quantifies the domain gap challenges in real-world re-id scenarios.

Conclusions:

• Standardized benchmarks and systematic evaluations are vital for reproducible re-id research.
• CLIP-based approaches show potential for improving model robustness and generalization in diverse surveillance applications.
• This work provides insights into mitigating domain gap issues in person re-identification.