Reproduce “Mitigating Resolution-Drift in Federated Learning: Case of Keypoint Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Taeheon Lim, Joohyung Lee, Kyungjae Lee, Jungchan Cho

Researchers at Chung-Ang University, Gachon University, and Korea University developed Resolution Adaptive Federated Learning (RAF) to mitigate "resolution-drift" in federated learning for high-resolution tasks like human pose estimation. RAF integrates a multi-resolution knowledge distillation strategy and modified Vision Transformers, achieving substantial accuracy improvements (e.g., 27.6% at 128x96) and improved generalization across diverse input resolutions.

Key insights

  • "Resolution-drift" is a distinct form of statistical heterogeneity in FL for high-resolution tasks, where aggregated global models suffer performance degradation due to clients' varying input resolutions.
  • Multi-resolution knowledge distillation effectively regularizes local models by enforcing scale-consistency across predictions, preventing overfitting to specific client resolutions and enhancing generalization.
  • Vision Transformers, with modified convolution-based positional embeddings, can dynamically adapt to and process arbitrary input resolutions in an FL setting, overcoming the limitations of standard fixed positional embeddings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.