Research map / 3D vision / Face and shape
ProbeSDF: Light Field Probes For Neural Surface 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.
Re-parametrizes SDF-based neural surface reconstruction appearance with decoupled spatial voxel features and coarse-grid spherical-harmonic light field probes fed to a tiny MLP, achieving faster training and better chamfer/PSNR.
Problem
Appearance models in SDF-based differentiable rendering for multiview 3D reconstruction co-locate spatial and angular features, which is costly in memory and training time and limits quality.
Approach
Decouple radiance into high-resolution spatial features (outer product of three factorized planes) and low-resolution spherical-harmonic light field probes indexed by the reflected vector, plus a polynomial Fresnel-like term in (1 - n·v), decoded by a tiny MLP fused into one CUDA kernel.
Key results
- On MVMannequins, the (4,4,4) configuration reaches 1.04 mm Chamfer and 36.81 PSNR in about 1 minute of training, versus 1.14 mm for MMH.
- On DTU, the (4,4,4) configuration with per-camera bias reaches 0.68 Chamfer and 37.03 PSNR in 150s, outperforming Voxurf, NeuS2 and 2DGS.
- On ActorsHQ, the method reaches 37.48 PSNR at full resolution in about 4 minutes, versus 36.56 for NeuS2 at half resolution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the baseline appearance model with ProbeSDF | MVMannequins reconstruction · MVMannequins | Chamfer (mm): 1.14 -> 1.04 | Helped |
| replace the baseline appearance model with ProbeSDF | ActorsHQ reconstruction · ActorsHQ | PSNR: 34.74 -> 37.48 | Helped |
| replace the baseline appearance model with ProbeSDF | DTU reconstruction · DTU | Chamfer: 0.73 -> 0.68 | Helped |
| replace the baseline appearance model with ProbeSDF | BlendedMVS reconstruction · BlendedMVS | PSNR: 35.11 -> 35.89 | Helped |
| remove probe smoothing (ablation) | ProbeSDF · MVMannequins | Chamfer (mm): 1.04 -> 1.09 | Hurt |
| remove probe smoothing (ablation) | ProbeSDF · MVMannequins | PSNR: 36.81 -> 36.91 | No clear effect |
| remove the Fresnel term (ablation) | ProbeSDF · MVMannequins | Chamfer (mm): 1.04 -> 1.18 | Hurt |
| remove the Fresnel term (ablation) | ProbeSDF · MVMannequins | PSNR: 36.81 -> 36.79 | No clear effect |
| add per-camera bias vectors (ablation) | ProbeSDF (4,4,4) · BlendedMVS | Chamfer: 2.47 -> 2.37 | Helped |
| add per-camera bias vectors (ablation) | ProbeSDF (4,4,4) · BlendedMVS | PSNR: 34.76 -> 35.19 | Helped |
| increase spatial and angular feature counts from (4,4,4) to (8,8,4) (ablation) | ProbeSDF with per-camera bias vectors · BlendedMVS | Chamfer: 2.37 -> 2.22 | Helped |
| increase spatial and angular feature counts from (4,4,4) to (8,8,4) (ablation) | ProbeSDF with per-camera bias vectors · BlendedMVS | PSNR: 35.19 -> 35.89 | Helped |
Limitations
- Locally encoded lighting degrades extrapolation compared with global environment representations, causing shadow artifacts on ActorsHQ legs.
- No explicit light transport simulation, so local minima from shape-radiance ambiguity can occur; relighting and material editing are unsupported.
- Assumes isotropic materials with mild-to-medium roughness and far lights; high-frequency reflections are not handled.
Experimental setup
Datasets: MVMannequins ActorsHQ DTU BlendedMVS
Benchmarks: MVMannequins ActorsHQ DTU BlendedMVS
Metrics: Chamfer distance PSNR training time rendering speed model size
Baselines: MMH Voxurf NeuS2 2DGS Colmap
Builds on: NeuS Voxurf NeuS2 Millimetric humans (MMH) 3D Gaussian Splatting
Closest papers by meaning
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.88)
- AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction (ICLR 2025 · similarity 0.88)
- HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces (CVPR 2024 · similarity 0.88)
- DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions (NeurIPS 2026 · similarity 0.88)
- LIRM: Large Inverse Rendering Model for Progressive Reconstruction of Shape, Materials and View-dependent Radiance Fields (CVPR 2025 · similarity 0.87)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 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.