- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a learning-guided pipeline for converting gridded SDFs into adaptive tetrahedral meshes. A 3D U-Net predicts surface and volume sampling densities, supervised by CGAL meshes. Seeds are drawn by GPU blue-noise sampling, Delaunay-tetrahedralized, and clipped with marching tetrahedra, and a classical post-processing stage then repairs dents, spikes, non-manifold elements, and slivers. The key idea is to replace per-shape iterative refinement with a feed-forward seed initialization. On 797 shapes from four datasets, the method matches CGAL’s accuracy while running about 2× faster, and produces closed, manifold, sliver-free meshes, at the cost of 1.4–1.6× more tetrahedra. On TripoSG-generated SDFs, it meshes every shape, whereas CGAL often drops parts or fails.
Critical strengths
-
The design cleanly separates learning from geometry: the network only predicts where to place seeds, while connectivity, boundary fitting, and element quality are handled by classical algorithms. This sidesteps the difficulty of learning discrete tetrahedral connectivity directly, and it yields practically valid output: every mesh is closed, manifold, and sliver-free, with zero failures across 2,787 cases. That matters for downstream simulation, where invalid meshes are unusable.
-
The method is practically relevant to generative 3D pipelines. It works directly on SDFs, needs no changes to the generation stage, and generalizes without fine-tuning to SDFs generated by TripoSG. On those, it meshes every shape, while CGAL frequently drops disconnected parts or times out. This shows a concrete use case where the approach is more robust than the standard tool.
Critical weaknesses
-
The learned component’s contribution is not isolated. The main novelty is a feed-forward seed initialization from a standard 3D U-Net, and the rest of the pipeline is classical. There is no comparison with the same pipeline using a non-learned density (uniform, curvature-based, or local feature size, as in Guo et al. 2016). Since post-processing takes about 70–75% of runtime and provides all validity guarantees, the gains may come mostly from blue-noise sampling and post-processing rather than learning. The model itself is also under-analyzed: its loss terms, branches, and KDE targets are not ablated, and density accuracy is never reported.
-
The evaluation setting limits the conclusions. The pipeline works on a rasterized SDF grid, so it is adaptive only in element size, and fidelity stays capped by the grid. Accuracy is measured against the marching-cubes surface of the same grid rather than the ground-truth meshes, which hides errors at sharp features. The closest baselines are missing: sampling plus Delaunay with a hand-designed sizing field, extraction plus remeshing plus tetrahedralization, and isosurface stuffing. Several comparisons are also unbalanced: GPU versus CPU, untuned CGAL parameters (CGAL is also the teacher), and TetGen and Gmsh forced to keep the dense marching-cubes surface.
Questions for authors
-
How does the same pipeline perform with a non-learned density (uniform, curvature-based, or local feature size), or with Poisson-disk sampling plus Delaunay as in Guo et al. (2016)? This is needed to isolate what learning contributes.
-
Please compare against an extract–remesh–tetrahedralize pipeline (e.g., dual contouring or FlexiCubes, adaptive remeshing, then TetGen or fTetWild) and against isosurface stuffing.
-
Please report accuracy against the original ground-truth meshes, including sharp-feature or Hausdorff metrics, not only against the marching-cubes surface of the input grid.
-
Why must the geometric stages (projection, bisection, surface recovery) query the rasterized grid rather than a continuous SDF, especially for neural fields such as TripoSG’s?
-
How do runtimes compare when baselines are hardware-matched, and how sensitive are CGAL’s speed and quality to its parameters? Can CGAL’s failures be avoided by supplying initial seed points?
-
Please give a per-stage runtime breakdown, time distributions (median and max), and per-dataset results, especially for the zero-shot sets and the generated SDFs.
-
How accurate is the predicted density itself, and which loss terms and branches (global, local, KDE targets) actually matter?
-
Whether density prediction is purely local and deterministic, and how well it generalizes?