Research map / 3D vision / Face and shape
GURecon: Learning Detailed 3D Geometric Uncertainties 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.
Proposes GURecon, which learns a continuous 3D geometric uncertainty field for neural SDF surfaces by online distillation of patch-based multi-view photometric consistency, with decoupled view-dependent fields to refine it.
Problem
Assessing geometric quality of neural surface reconstructions without ground-truth meshes; existing NeRF uncertainty methods are pixel-wise and rendering-based, so they do not capture view-independent 3D geometric uncertainty.
Approach
Use patch-based homography-warped multi-view consistency (SSIM) at zero-crossing surface points as pseudo labels to distill a continuous 3D uncertainty field online, and learn decoupled view-independent/view-dependent fields to remove lighting effects before recomputing consistency for fine-tuning.
Key results
- On BlendedMVS, the full model reaches AUSE_MSE 0.058, AUSE_MAE 0.094 and AUSE_3D 0.091 in the ablation table, outperforming the ablated variants.
- On TNT scenes, GURecon attains AUSE_3D of 0.327 (Barn), 0.243 (Truck) and 0.198 (Caterpillar), versus 0.438, 0.289 and 0.346 for Bayes' Rays.
- In the incremental NBV experiment on TNT with 30% of views, GURecon gives the lowest Chamfer Distance on Barn (0.947), Caterpillar (0.705) and Truck (2.059), and the highest PSNR on each scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GURecon instead of Bayes' Rays | hash-based NeuS · TNT-Barn | AUSE_3D: 0.438 -> 0.327 | Helped |
| use GURecon instead of Bayes' Rays | hash-based NeuS · TNT-Truck | AUSE_3D: 0.289 -> 0.243 | Helped |
| use GURecon instead of Bayes' Rays | hash-based NeuS · TNT-Caterpillar | AUSE_3D: 0.346 -> 0.198 | Helped |
| use GURecon for next-best-view selection instead of random selection | incremental reconstruction · TNT-Barn | CD: 1.033 -> 0.947 | Helped |
| use GURecon for next-best-view selection instead of random selection | incremental reconstruction · TNT-Caterpillar | CD: 0.809 -> 0.705 | Helped |
| use GURecon for next-best-view selection instead of random selection | incremental reconstruction · TNT-Truck | CD: 2.248 -> 2.059 | Helped |
| add decoupled-field finetuning (ablation) | GURecon · BlendedMVS | AUSE_MSE: 0.104 -> 0.058 | Helped |
| add decoupled-field finetuning (ablation) | GURecon · BlendedMVS | AUSE_3D: 0.125 -> 0.091 | Helped |
| replace patch consistency with pixel-based consistency (ablation) | GURecon · BlendedMVS | AUSE_3D: 0.091 -> 0.792 | Hurt |
| use smaller patch size 7 instead of 11x11 (ablation) | GURecon · BlendedMVS | AUSE_3D: 0.091 -> 0.179 | Hurt |
| use larger patch size 15 instead of 11x11 (ablation) | GURecon · BlendedMVS | AUSE_3D: 0.091 -> 0.214 | Hurt |
Surprising findings
- Pixel-based consistency performs much worse than patch-based consistency (AUSE_3D 0.792 vs 0.091), showing sensitivity to lighting variation.
Limitations
- Performance is limited in extreme scenarios with large textureless areas such as white walls.
- Bayes' Rays and ActiveNeRF are only approximately applied to surface points, so 3D comparisons are approximate.
- Relies on structural modifications of NeRF baselines to work on SDF backends.
Experimental setup
Datasets: DTU BlendedMVS Tanks and Temples (TNT)
Benchmarks: DTU BlendedMVS Tanks and Temples
Metrics: AUSE (MSE, MAE, 3D) Chamfer Distance PSNR
Baselines: CFNeRF ActiveNeRF Bayes' Rays Uncertainty-Guided NeRF Lee et al.
Builds on: NeuS Instant-NGP Bayes' Rays ActiveNeRF 2D Gaussian Splatting
Closest papers by meaning
- PhyRecon: Physically Plausible Neural Scene Reconstruction (NeurIPS 2024 · similarity 0.87)
- UFORecon: Generalizable Sparse-View Surface Reconstruction from Arbitrary and Unfavorable Sets (CVPR 2024 · similarity 0.87)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.86)
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors (ICLR 2024 · similarity 0.86)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.86)
- AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction (ICLR 2025 · 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.