Person re-identification based on multi-branch visual transformer and self-distillation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-ID) technology has advanced with deep learning.
• Current re-ID models exhibit lower recognition rates than face recognition, limiting practical applications.
• Enhancing pedestrian re-identification model accuracy remains a critical research objective.

Purpose of the Study:

• To improve the recognition rate of person re-identification models.
• To explore effective online self-distillation strategies within multi-branch networks.
• To analyze knowledge distillation methods and the impact of noise perturbation on model performance.

Main Methods:

• Utilized multi-branch network characteristics for efficient online self-distillation without extra resources.
• Theoretically and experimentally analyzed knowledge distillation using Mean Squared Error (MSE) and Kullback-Leibler (KL) divergence.
• Introduced specific noise perturbation to model weights to boost recognition accuracy.

Main Results:

• Developed an effective online self-distillation scheme leveraging multi-branch network information.
• Provided a comparative analysis of MSE and KL divergence for knowledge distillation in re-ID.
• Demonstrated that noise perturbation significantly improves model recognition rates.
• Achieved state-of-the-art performance on four public person re-identification datasets.

Conclusions:

• Optimized multi-branch networks through effective self-distillation and noise perturbation enhance person re-identification.
• The study provides valuable insights into knowledge distillation techniques for re-ID tasks.
• The proposed methods collectively yield state-of-the-art results, advancing practical person re-identification applications.