University of Trento's TetraCoder Ditches Hash Grids for Exact 3D Mesh Extraction

A new neural field encoder replaces multiresolution hash grids with a single adaptive tetrahedral mesh, enabling exact mesh extraction and editing.

·
·
·
University of Trento's TetraCoder Ditches Hash Grids for Exact 3D Mesh ExtractionPRO
Read2 min
TypePaper
TopicThree D · Image
  • TetraCoder replaces multiresolution hash grids with a single adaptive tetrahedral mesh for neural fields.
  • The encoder plus decoder is continuous piecewise affine, giving exact analytic mesh extraction from the learned SDF.
  • No hash collisions, since each vertex owns its own feature vector in a one-to-one mapping.
  • Scales better than multiresolution encoders and uses fewer parameters at matched quality.
  • Tetrahedral scaffold doubles as a ray-tracing accelerator and supports direct geometry and appearance editing.
  • Beats state of the art on 2D/3D fitting, novel view synthesis, reconstruction, and editing benchmarks.

University of Trento researchers have introduced TetraCoder, a single-level, hash-free spatial encoder built on an adaptive tetrahedral mesh. The representation keeps a learned signed distance field, or SDF, in exact correspondence with an analytic polygonal surface when paired with a continuous piecewise-affine decoder.

Neural rendering and reconstruction pipelines commonly use multiresolution hash grids to encode spatial coordinates. Those grids train quickly and use memory efficiently, but hash collisions force unrelated locations to share features. Their trilinear interpolation also complicates exact mesh extraction, leaving most systems to sample the SDF on a dense grid and run marching cubes.

Comparison of multiresolution hash grids, tetrahedral hash grids, and TetraCoder's adaptive tetrahedral mesh
TetraCoder replaces stacked hashed grids with one adaptive tetrahedral mesh.

Where hash grids lose exactness

Multiresolution hash encoders store learnable feature vectors at grid vertices across several resolutions. A query point retrieves features through trilinear interpolation at every level, and the resulting vectors feed a neural decoder that predicts density, color, or signed distance.

Trilinear interpolation is continuous inside a voxel, but it is not affine because it contains products of spatial coordinates. A continuous piecewise-affine function has planar level sets within each linear region, so its zero level set can be recovered as polygonal facets. Trilinear terms remove that direct correspondence and usually require sampled extraction.

Hashing introduces a separate source of ambiguity. Multiple grid vertices can map to the same table entry, leaving the decoder to distinguish spatially unrelated points from surrounding context. TetraCoder assigns features directly to mesh vertices, giving each stored feature a unique spatial location.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Trending
  • No trending articles

Comments

avatar

Next Reads