Reproduce “DSAC - Differentiable RANSAC for Camera Localization” 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

Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, Carsten Rother

Researchers at TU Dresden and Microsoft developed DSAC, a differentiable variant of RANSAC that enables end-to-end training of robust geometric algorithms within deep learning pipelines. DSAC achieved a new state-of-the-art accuracy of 62.5% on the 7-Scenes dataset for RGB-only camera localization, significantly outperforming prior methods.

Key insights

  • Probabilistic hypothesis selection (DSAC) effectively preserves the "hard decision" principle of RANSAC while enabling end-to-end differentiability, unlike "soft argmax" averaging which degrades performance.
  • End-to-end training with DSAC leads to "cautious" and widespread improvements in learned scene coordinates, resulting in stable and robust learning across the network.
  • DSAC maintains a higher entropy in its score distribution compared to SoftAM, allowing for better probabilistic exploration of hypotheses and preventing overfitting to overly confident, potentially incorrect, selections.

Summary by alphaXiv.

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