Reproduce “CrossSDF: 3D Reconstruction of Thin Structures From Cross-Sections” 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

Thomas Walker, Salvatore Esposito, Daniel Rebain, Amir Vaxman, Arno Onken, Changjian Li, Oisin Mac Aodha

CrossSDF reconstructs continuous 3D surfaces from planar contours by learning a signed distance field. It uses slice-derived distances at contours and where inside/outside predictions disagree, with hybrid hash-grid/Fourier features. On 100 aligned Alveolis sections, its Chamfer Distance is 0.35 versus OReX’s 3.6 (both ×100), but it yields 6 connected components versus 1 in the reference; very sparse slices remain difficult.

Key insights

  • Direct regression to 2D SDF values can lead to "laddering" artifacts; a "contour-aware" loss focused on classification disagreement resolves this while maintaining fidelity.
  • Adaptive sampling inside contours is crucial for faithfully reconstructing thin structures often undersampled by uniform or fixed-radius methods.
  • Combining hash-grid encodings and Random Fourier Features effectively captures fine details while maintaining surface smoothness and reducing grid interpolation artifacts in sparse data.

Summary by alphaXiv.

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