Dual-Refinement: Joint Label and Feature Refinement for Unsupervised Domain Adaptive Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptive (UDA) person re-identification (re-ID) faces challenges due to the absence of target domain labels.
• Existing methods use off-line clustering for pseudo-labels, but these often contain noise, limiting performance.
• Noisy pseudo-labels hinder the discriminability of learned features in the target domain.

Purpose of the Study:

• To propose a novel approach, Dual-Refinement, for improving UDA person re-ID.
• To jointly refine pseudo-labels and features to enhance label purity and feature discriminability.
• To overcome the limitations of noisy pseudo-labels in existing UDA re-ID methods.

Main Methods:

• A hierarchical clustering scheme is introduced for off-line pseudo-label refinement using representative prototypes.
• An instant memory spread-out (IM-spread-out) regularization is proposed for on-line feature learning.
• An instant memory bank stores sample features for spread-out learning across the entire dataset.

Main Results:

• The Dual-Refinement method effectively reduces the impact of noisy labels.
• Learned features are refined through an alternative training process, boosting discriminability.
• Experimental results show significant outperformance compared to state-of-the-art methods.

Conclusions:

• Dual-Refinement offers a robust solution for UDA person re-ID by addressing noisy pseudo-labels.
• The joint refinement of labels and features leads to more reliable re-identification.
• This approach advances the field of unsupervised domain adaptive person re-identification.