MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement

2 min read

MoGe-3 targets the over-smoothed thin structures and small objects that monocular geometry models still get wrong. The authors blame decoding 3D geometry in a 2D image grid, where features mix by image-plane proximity, so surfaces far apart in depth bleed into each other. Their fix lifts the coarse point map from a base model onto a sparse voxel shell and refines it with sparse 3D convolutions (Self-Guided Sparse 3D Refinement), so features only mix with neighbours that are close in 3D.

Findings

  • The abstract reports clear gains over existing approaches on fine 3D detail across several datasets, in metrics and visually. No headline number in the abstract.
  • The README notes it depends on FlexGEMM, built on Triton, so it does not run on macOS.

Limits

  • I read the abstract and README only.
  • Refinement is only as good as the coarse map it starts from, and it adds a sparse 3D stage to the cost. Nothing in what I read covers scattering media.

Relevance

Thin, depth-discontinuous structure is where monocular models fail in ways that affect a robot: a smeared pole or cable is an obstacle in the wrong place. It is worth testing whether 3D-local refinement also stops haze or fog from smearing foreground objects into the background. See foundation models for robotics and MoGe-2.