Reproduce “BMD: A General Class-balanced Multicentric Dynamic Prototype Strategy for Source-free Domain Adaptation” on OpenResearch

About this paper

Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, Dacheng Tao

Researchers from Tongji University, The University of Sydney, and other institutions developed the Class-balanced Multicentric Dynamic (BMD) prototype strategy for Source-free Domain Adaptation (SFDA). The approach, designed to improve pseudo-label quality, consistently boosted the accuracy of various SFDA methods on 2D image and 3D point cloud datasets, achieving an 88.7% average accuracy with SHOT++ on VisDA-C and demonstrating more balanced performance across challenging classes.

Key insights

• Globally selecting top-scoring instances for each class rather than relying on local predictions can effectively balance prototype generation across categories, which is crucial for robust pseudo-labeling in SFDA.
• Representing diverse categories with multiple, fine-grained prototypes (multicentric) provides a more accurate and nuanced understanding of class feature distributions, improving pseudo-label precision for hard-to-adapt instances.
• Integrating dynamic, real-time updates from the network's optimization process via Exponentially Moving Average (EMA) with more stable, epoch-level updates leads to a more responsive and robust adaptation process, addressing the limitations of static pseudo-labeling.

Summary by alphaXiv.