Research map / Point clouds / Shape and surface completion
Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors
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.
Finetunes a data-driven DeepSDF-style neural SDF prior on a single noisy point cloud using a local statistical noise-to-noise mapping loss (EMD between pulled queries and noisy points in KNN local regions), without SDF supervision, clean points, or normals.
Problem
Estimating accurate neural signed distance functions from highly noisy point clouds, where data-driven priors generalize poorly and overfitting-based methods converge slowly.
Approach
Initialize an overfitting stage from a data-driven SDF prior (auto-decoder trained with SDF supervision) and finetune it on a single noisy point cloud with a local noise-to-noise mapping loss: pulled query projections are matched by EMD to random subsets of noisy points in K-nearest-neighbor local regions.
Key results
- On ShapeNet, achieves CD_L1 x10 of 0.023 vs 0.026 for N2NM (prior best), NC 0.973 and F-Score 0.992.
- On ABC middle noise, CD_L2 x100 of 0.096 vs 0.113 for N2NM; test-time overfitting takes about 5 min vs 46 min for N2NM on ShapeNet.
- On 3D Scene, CD_L1 0.016 with NC 0.942, outperforming N2NM (CD_L1 0.019, NC 0.929) and LIG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| finetune a data-driven prior using local noise-to-noise mapping | Neural SDF inference from noisy point clouds · ShapeNet | CD_L1 ×10: 0.026 -> 0.023 | Helped |
| finetune a data-driven prior using local noise-to-noise mapping | Neural SDF inference from noisy point clouds · ABC var and ABC max | CD_L2 ×100: 0.113 -> 0.096; 0.139 -> 0.113 | Helped |
| finetune a data-driven prior using local noise-to-noise mapping | Neural SDF inference from noisy point clouds · SRB | CD_L1 ×10: 0.067 -> 0.055 | Helped |
| finetune a data-driven prior using local noise-to-noise mapping | Neural SDF inference from noisy point clouds · 3D Scene | CD_L1: 0.019 -> 0.016 | Helped |
| use embedding size 256 rather than 128 or 512 (ablation) | Proposed method · ABC | CD_L2 ×100: 0.102 -> 0.096; 0.114 -> 0.096 | Helped |
| add the learned prior and learned embedding initialization (ablation) | Proposed method · ABC | CD_L2 ×100: 0.108 -> 0.096 | Helped |
| optimize the learned network parameters rather than fixing them (ablation) | Proposed method · ABC | CD_L2 ×100: 0.144 -> 0.096 | Helped |
| use KNN sphere regions rather than voxel or fixed-radius regions (ablation) | Proposed method · ABC | CD_L2 ×100: 0.314 -> 0.096; 0.101 -> 0.096 | Helped |
| use local rather than global noise-to-noise mapping (ablation) | Proposed method · ABC | CD_L2 ×100 and time: 0.106 -> 0.096; 21 min -> 8 min | Helped |
| use 1000 points per local region rather than 500, 3000, or 5000 (ablation) | Proposed method · ABC | CD_L2 ×100: 0.102 -> 0.096; 0.111 -> 0.096; 0.114 -> 0.096 | Helped |
| initialize the SDF from the learned prior rather than random, square, or Sphere (SAL) initialization (ablation) | Proposed method · ABC | Time: 8.3min -> 5.0min; 7.1min -> 5.0min; 5.5min -> 5.0min | Helped |
| increase noise from middle to max or extreme (ablation) | Ours · ABC | CD_L2: 0.096 -> 0.113 -> 0.125 | Hurt |
Surprising findings
- Fixing the learned prior parameters and optimizing only the embedding (auto-decoding style) degrades results (CD_L2 0.144 vs 0.096).
Limitations
- Fails when noise is too large to corrupt the shape heavily, per authors' appendix.
- Prior is trained on ShapeNet and transferred to scans and scenes without retraining, since there are limited scene training samples.
Experimental setup
Datasets: ShapeNet ABC FAMOUS SRB D-FAUST 3D Scene KITTI Paris-rue-Madame nuScenes
Benchmarks: ShapeNet ABC SRB FAMOUS D-FAUST 3D Scene KITTI Paris-rue-Madame
Metrics: CD_L1 CD_L2 NC F-Score inference time
Baselines: PSR PSG R2N2 COcc SAP OCNN IMLS POCO ALTO N2NM P2S NP
Builds on: DeepSDF Noise2NoiseMapping NeuralPull ALTO
Closest papers by meaning
- Toward Robust Neural Reconstruction from Sparse Point Sets (CVPR 2025 · similarity 0.89)
- Unsupervised Occupancy Learning from Sparse Point Cloud (CVPR 2024 · similarity 0.88)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.87)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.87)
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (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.