[fVDB Reality Capture] Poor mesh quality from 3DGS — parameter tuning & alternatives

Omniverse Core Platform NuRec
1

Hello,

Environment: fvdb-reality-capture (latest), spatial video → 3DGS → tsdf_from_splats_dlnr → mesh

Problem: The output mesh looks as if a small mesh is being generated around each individual Gaussian splat rather than a clean, continuous surface.

Q1 — Alternative or upcoming APIs for 3DGS-to-mesh?

I’m already aware of mesh_from_splats_dlnr(). Are there any other NVIDIA-maintained tools or pipelines for higher-quality mesh extraction from 3DGS — including anything currently in beta or active development (e.g., within the NVIDIA Spatial Intelligence Lab)?

Q2 — Parameter tuning for tsdf_from_splats_dlnr

Which parameters have the most impact on mesh quality for indoor room-scale scenes?

  • voxel_size: recommended range?
  • truncation: should it be a fixed multiple of voxel_size (e.g., 3×)?
  • min_weight or similar: is there a way to discard voxels with too few observations?
  • Marching cubes isolevel: is this exposed, and does it help with surface smoothness?

Q3 — Removing floating / mid-air noise Gaussians

I have many floating Gaussians in free-air regions that propagate into the TSDF as noisy depth estimates. Is there a built-in opacity- or density-based pruning step in fvdb-reality-capture, or is transforms.CropScene / CropSceneToPoints the intended approach for this?

Any guidance from the team or community would be appreciated. Thank you.

image
image1202×842 245 KB

Thanks

6.0.0
5.1.0
5.0.0
4.5.0
4.2.0
4.1.0
4.0.0
4.5.0
2023.1.1
2023.1.0-hotfix.1
Other (please specify):

Operating System

Ubuntu 24.04
Ubuntu 22.04
Ubuntu 20.04
Windows 11
Windows 10
Other (please specify):

GPU Information

  • Model: Blackwell Max-Q

  • Driver Version:

2

We ran a side-by-side comparison on tsdf_from_splats_dlnr with two configs that reproduce exactly the “mesh-around-each-splat” artifact you’re seeing vs. a clean surface. Results below.

Comparison configs (mesh_compare.py)

Param “Bad” (noisy, splat-ish) “Good” (clean surface)
voxel_size 0.005 0.02
truncation 1.0 × voxel (=0.005) 4.0 × voxel (=0.08)
min_weight 1 3
prune_opacity None 0.1

Q2 — Parameter tuning

For indoor room-scale scenes the dominant factors, in order of impact:

  1. voxel_size — counter-intuitive, but going coarser helps a lot. 0.005 m tries to resolve detail finer than the splat noise floor and ends up carving a tiny isosurface around every Gaussian. 0.01–0.03 m is the sweet spot for room-scale; we used 0.02.

  2. truncation (a.k.a. truncation margin) — should be a multiple of voxel_size, not equal to it. A 1× truncation gives each splat its own little TSDF bubble. 3–4× voxel_size lets neighbouring splats’ SDFs blend into a single continuous surface. We used 4×.

  3. min_weight — raise it. min_weight=1 keeps every voxel a single splat touched (= floaters survive). min_weight=3 drops voxels that weren’t confirmed by multiple views/splats and removes most mid-air noise.

  4. Marching-cubes isolevel — not currently exposed as a tunable in tsdf_from_splats_dlnr; in practice the three params above dominate, so we didn’t need it.

So yes — truncation = ~4 × voxel_size is a good rule of thumb, and min_weight ≥ 3 is what you want to try.

Q3 — Removing floating / mid-air Gaussians

There is a built-in pruning step: prune_opacity in the TSDF call. Setting prune_opacity=0.1 drops low-opacity Gaussians before they contribute to the SDF, which kills most of the airborne floaters at the source. This was the single biggest visual win in the comparison.

comparison
comparison1393×760 116 KB

.