Reproduce “FREST: Feature RESToration for Semantic Segmentation under Multiple Adverse Conditions” 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

Sohyun Lee, Namyup Kim, Sungyeon Kim, Suha Kwak

A new framework, FREST, enhances semantic segmentation under diverse adverse conditions by restoring features of degraded images to simulate normal conditions in a source-free domain adaptation setting. It achieves a 70.7% mIoU on the ACDC benchmark, outperforming prior state-of-the-art methods.

Key insights

  • Instead of forcing full condition-invariance, restoring only adverse image features to simulate normal conditions effectively preserves valuable source domain knowledge, mitigating catastrophic forgetting.
  • A dedicated, lightweight "condition strainer" module can effectively extract and embed condition-specific information without distorting semantic content or adding significant inference overhead.
  • Leveraging a condition embedding space and a contrastive loss that focuses on condition similarity (rather than pixel-wise semantic similarity) makes the adaptation robust to content misalignments inherent in real-world paired data.

Summary by alphaXiv.

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