Reproduce “SMD-Nets: Stereo Mixture Density Networks” 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

Fabio Tosi, Yiyi Liao, Carolin Schmitt, Andreas Geiger

SMD-Nets represents disparity at continuous image locations with a two-mode Laplacian mixture, allowing it to choose between foreground and background disparities at boundaries. In a PSM ablation on UnrealStereo4K, this output lowers SEE₃ to 1.52 from 1.98 for point regression and 2.50 for a unimodal output, while EPE is 1.11, close to point regression’s 1.10. These results concern boundary error, not downstream performance.

Key insights

  • Modeling disparity as a bimodal mixture distribution effectively captures distinct foreground and background disparities at object boundaries, leading to sharp, discontinuous transitions.
  • A continuous output representation enables efficient, arbitrary-resolution disparity estimation (stereo super-resolution) without requiring architectural modifications.
  • Targeted training on depth discontinuities using DDA sampling significantly improves a model's ability to resolve ambiguities and generate accurate boundaries.

Summary by alphaXiv.

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