Research map / 3D vision / NeRF and radiance fields
AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction
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.
Proposes AniSDF, an SDF-based neural surface method with parallel coarse/fine hash grids (fused granularity) and blended view/reflection radiance fields using anisotropic spherical Gaussian encoding, for joint high-fidelity geometry and rendering.
Problem
SDF-based neural surface methods struggle to jointly model fine geometric details and disambiguate geometry from reflective/complex appearance, trading off geometry accuracy against rendering quality.
Approach
Train coarse and fine multi-resolution hash grids in parallel (fused granularity) rather than sequentially coarse-to-fine, and split appearance into a view-based diffuse radiance field and a reflection-based radiance field encoded with anisotropic spherical Gaussians, blended by a learned 3D weight field.
Key results
- On NeRF Synthetic, AniSDF reaches 33.42 average PSNR and 4.89e-3 average Chamfer Distance, the best on both among compared methods.
- On Shiny Blender, AniSDF averages 31.05 PSNR and 0.85 MAE, outperforming RefNeuS (30.20 PSNR, 0.91 MAE).
- Ablation on NeRF Synthetic: removing both ASG and fused granularity drops PSNR from 33.42 to 30.25 and raises CD from 4.89 to 5.64.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ASG encoding (ablation) | AniSDF · NeRF Synthetic | PSNR: 32.38 -> 33.42 | Helped |
| add ASG encoding (ablation) | AniSDF · NeRF Synthetic | Chamfer Distance: 5.16 -> 4.89 | Helped |
| add fused-granularity surfaces (ablation) | AniSDF · NeRF Synthetic | PSNR: 33.19 -> 33.42 | Helped |
| add fused-granularity surfaces (ablation) | AniSDF · NeRF Synthetic | Chamfer Distance: 5.37 -> 4.89 | Helped |
| add both ASG encoding and fused-granularity surfaces (ablation) | AniSDF · NeRF Synthetic | PSNR: 30.25 -> 33.42 | Helped |
| add both ASG encoding and fused-granularity surfaces (ablation) | AniSDF · NeRF Synthetic | Chamfer Distance: 5.64 -> 4.89 | Helped |
| use maximum hashgrid resolution 1024 instead of the default 2048 (ablation) | AniSDF · NeRF Synthetic ship | geometric detail reconstruction | Hurt |
| use maximum hashgrid resolution 4096 instead of the default 2048 (ablation) | AniSDF · NeRF Synthetic ship | geometric detail reconstruction | Helped |
| replace fused granularity with two fine branches (ablation) | AniSDF · NeRF Synthetic ship | surface quality | Hurt |
| replace fused granularity with two coarse branches (ablation) | AniSDF · NeRF Synthetic ship | geometric detail reconstruction | Hurt |
| set granularity boundary level m to 9 or 11 instead of the default (ablation) | AniSDF · NeRF Synthetic lego | reconstruction quality | Hurt |
Surprising findings
- Chamfer distance does not always reflect geometry quality for hollow objects like mic, per the authors' note that solid-geometry predictions can score better numerically.
Limitations
- Cannot achieve real-time rendering.
- Fails in scenes with complex indirect illumination due to lack of a materials estimation network.
- Chamfer distance may not reliably reflect geometry quality for hollow objects.
Experimental setup
Datasets: NeRF Synthetic DTU Shiny Blender Shelly MipNeRF360 NeRO Glossy-Real
Benchmarks: NeRF Synthetic Shiny Blender DTU MipNeRF360 Shelly
Metrics: PSNR Chamfer Distance MAE
Baselines: NeRF InstantNGP Mip-NeRF Zip-NeRF 3DGS NeuS NeRO BakedSDF NeRF2Mesh 2DGS RefNeuS Neuralangelo
Builds on: NeuS UniSDF PermutoSDF RefNeRF Anisotropic spherical Gaussians
Closest papers by meaning
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.89)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.88)
- GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction (NeurIPS 2024 · similarity 0.88)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.88)
- DIFIX3D+: Improving 3D Reconstructions with Single-Step Diffusion Models (CVPR 2025 · similarity 0.87)
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.87)
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.