Reproduce “Inverse Rendering for Modeling with Line Primitives” 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

Kenji Tojo, Ariel Shamir, Nobuyuki Umetani, Bernd Bickel

The authors reconstruct fuzzy objects from multi-view images as explicit polylines whose filtered, subpixel coverage produces a fuzzy appearance. A stochastic differentiable rasterizer optimizes line positions and opacity, while periodic updates adapt connectivity. On six synthetic Shelly scenes, average LPIPS is 0.046 versus 0.057 for 3D Gaussian Splatting, though its PSNR is lower; topology is not guaranteed physically accurate.

Key insights

  • Explicit 1D line primitives can effectively represent fuzzy, anisotropic object appearances, providing geometric information that volumetric methods lack for downstream applications like physical simulation and advanced shading.
  • A stochastic differentiable rasterizer, carefully designed for subpixel-width lines, can overcome gradient blocking from discrete depth testing, enabling optimization of dense, fuzzy line geometry.
  • Combining continuous differentiable rendering with discrete topological updates through adaptive re-lining is crucial for robust and efficient reconstruction of complex curve networks from image observations.

Summary by alphaXiv.

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