Reproduce “A simple, strong baseline for building damage detection on the xBD dataset” 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

Sebastian Gerard, Paul Borne-Pons, Josephine Sullivan

KTH Royal Institute of Technology researchers developed a simplified yet effective baseline for building damage detection on the xBD dataset, achieving performance comparable to complex competition winners while also uncovering a significant generalization failure of these models on entirely unseen disaster events.

Key insights

  • A significantly simpler model can achieve adequate performance (within 2 percentage points) compared to highly complex, ensemble-based state-of-the-art solutions, making advanced HADR AI more accessible.
  • Current deep learning models for building damage detection exhibit severe generalization failure when applied to truly novel disaster events, with performance on 'minor' and 'major' damage categories dropping by approximately 20% and 50% respectively.
  • Dataset characteristics, such as highly unequal class distributions across different disaster events and the variability in damage appearance, contribute substantially to the observed generalization difficulties.

Summary by alphaXiv.

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