Reproduce “Robust Monocular Depth Estimation under Challenging 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

Stefano Gasperini, Nils Morbitzer, HyunJun Jung, Nassir Navab, Federico Tombari

This research from the Technical University of Munich and Google proposes `md4all`, a training technique for monocular depth estimation that improves robustness in challenging conditions like night and rain. By translating ideal condition images into synthetic adverse ones for training while using ideal condition ground truth, the method reduces depth estimation errors by 40-50% at night and 30-35% in rain.

Key insights

  • Image translation models can synthesize adverse conditions while retaining the underlying depth structure of ideal condition images.
  • Decoupling the visual input from the ground truth supervision signal allows models to learn robust depth features from challenging appearances without corrupting the learning process.
  • Applying this training methodology improves robustness across various base architectures and both self-supervised and fully-supervised depth estimation settings.

Summary by alphaXiv.

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