Research map / 3D vision / Face and shape
Volumetric Surfaces: Representing Fuzzy Geometries with Layered Meshes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Represents fuzzy geometry as k learned, semi-transparent SDF shells baked into lightweight meshes with neural SH textures, enabling sorting-free, few-sample rasterized real-time view synthesis on phones and laptops.
Problem
Real-time view synthesis of fuzzy geometry (hair, plush) on low-power devices: volume rendering needs many samples per ray, splatting needs sorting, and single-surface methods cannot model semi-transparent fuzzy structures.
Approach
Model k adaptively spaced SDF shells with a main surface plus offset support surfaces so layers are traversed in fixed order, add view-dependent transparency, bake each shell to a low-poly mesh with UV-based neural SH textures, and rasterize with fixed-order alpha blending.
Key results
- On Shelly, 7-Mesh reaches 34.50 PSNR versus 35.44 for 3DGS and 29.30 for MobileNeRF, at 42 FPS on a Samsung A52s smartphone versus 8 FPS for 3DGS.
- 3-Mesh reaches 65 FPS on the smartphone and 145 FPS on the Dell XPS 13 laptop, while 7-Mesh uses about 110 MB on disk versus 57 MB for 3DGS.
- Ablations show that removing curvature loss, the grazing-angle attenuation, or view-dependent transparency degrades the baked 5-Mesh results, and fixed offsets drop PSNR to 30.09.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove view-dependent surface transparency (ablation) | 5-Mesh · Shelly [56] | PSNR: 34.25 -> 32.71 | Hurt |
| remove curvature loss (ablation) | 5-Mesh · Shelly [56] | PSNR: 34.25 -> 33.41 | Hurt |
| remove transparency attenuation at grazing angles (ablation) | 5-Mesh · Shelly [56] | PSNR: 34.25 -> 33.96 | Hurt |
| use fixed support-surface offsets instead of trained offsets (ablation) | 5-Mesh · Shelly [56] | PSNR: 34.25 -> 30.09 | Hurt |
| initialize support surfaces outside the main SDF (ablation) | 5-Mesh · Shelly [56] | PSNR: 34.25 -> 30.85 | Hurt |
| increase the number of meshes from 7 to 9 | Volumetric Surfaces · Shelly | PSNR: 34.50 -> 34.38 | Hurt |
| increase the number of meshes from 7 to 9 | Volumetric Surfaces · Custom (plushy + [47]) | PSNR: 35.63 -> 35.74 | Helped |
| increase the number of meshes from 7 to 9 | Volumetric Surfaces · DTU scans 83, 105 | PSNR: 36.77 -> 37.17 | Helped |
| increase neural texture resolution up to 1024² | Volumetric Surfaces · khady scene from Shelly [56] | PSNR: 29.65 -> 30.02 | Helped |
| use mixed-resolution textures instead of all textures at 2048² | Volumetric Surfaces · khady scene from Shelly [56] | PSNR: 30.02 -> 30.02 | No clear effect |
Surprising findings
- Nine meshes give lower quality than seven under the same training iterations, because deeper shells receive smaller gradients.
- The implicit 5-SDF scores lower than its baked 5-Mesh counterpart, which the authors attribute to stochastic ray sampling spreading capacity off-surface.
Limitations
- Textured shells show artifacts at grazing angles, especially with limited capacity or sparse views.
- Performs poorly on sparsely observed scenes, explaining observations via view dependency rather than geometry.
- Handling thin structures is difficult due to the SDF geometry representation; gains on solid surfaces are marginal.
Experimental setup
Datasets: Shelly DTU QuadFields custom scenes (plushy, hairy monkey)
Benchmarks: Shelly DTU Custom plushy and hairy monkey scenes
Metrics: PSNR SSIM LPIPS FPS memory footprint (MB)
Baselines: 3DGS MobileNeRF Instant-NGP PermutoSDF AdaptiveShells QuadFields BakedSDF BOG
Builds on: NeuS PermutoSDF Textured shells MERF AdaptiveShells
Closest papers by meaning
- Geometry Field Splatting with Gaussian Surfels (CVPR 2025 · similarity 0.87)
- LinPrim: Linear Primitives for Differentiable Volumetric Rendering (NeurIPS 2025 · similarity 0.87)
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.87)
- FruitNinja: 3D Object Interior Texture Generation with Gaussian Splatting (CVPR 2025 · similarity 0.87)
- Ghost on the Shell: An Expressive Representation of General 3D Shapes (ICLR 2024 · similarity 0.86)
- HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.