Reproduce “Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models” 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

Georges Le Bellier, Nicolas Audebert

Researchers at Cnam and Univ. Gustave Eiffel developed ODEED, a novel diffusion-based scorer, to detect Out-Of-Distribution (OOD) Earth Observation images without supervision, particularly excelling in identifying flood-related anomalies. The method demonstrated superior performance in challenging 'near-OOD' scenarios like pre/post-flood and non-flooded/flooded imagery, achieving significant AUC and FPR95% improvements over autoencoder and discriminative baselines.

Key insights

  • Diffusion models effectively model the underlying data distribution of EO images and can be successfully applied for unsupervised OOD detection across diverse scenarios, including cloud detection and geographical shifts.
  • The deterministic reconstruction property of PF-ODE-based diffusion models offers a more stable and accurate OOD score (ODEED) compared to stochastic sampling approaches, proving particularly effective for 'near-OOD' detection.
  • The choice of reconstruction metric (MSE vs. LPIPS) and the corruption time hyperparameter ($t_0$) can significantly impact OOD detection performance, with MSE often better for localized anomalies like floods and LPIPS for global changes.

Summary by alphaXiv.

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