Research map / Point clouds / Shape and surface completion
Learning Signed Distance Functions from Noisy 3D Point Clouds via Noise to Noise Mapping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometric consistency regularization (lambda=0.1) to EMD noise-to-noise loss (ablation) | surface reconstruction · PU test set | L2CD x10^4 (reconstruction): 80.917 -> 4.277 | Helped |
| replace EMD distance with Chamfer distance in the noise-to-noise loss (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.221 -> 73.786 | Hurt |
| increase geometric consistency weight lambda from 0.1 to 0.2 (ablation) | denoising and reconstruction · PU test set | L2CD x10^4: denoise 4.252 -> 4.832 | Hurt |
| add geometric consistency (lambda 0.05 to 0.1) to EMD denoising (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.221 (λ=0) vs 4.245 (λ=0.05) vs 4.252 (λ=0.1) | No clear effect |
| increase number of noisy observations N from 1 to 200 (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.976 -> 4.221 | Helped |
| increase batch size from 250 to 10000 queries (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.221 -> 6.217 (B=10000); 12.398 at B=100 | Hurt |
| increase training iterations from 40 to 80 x10^4 (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.887 -> 4.221 | Helped |
| train on single noisy point cloud (N=1) instead of multiple observations (ablation) | point cloud denoising · PU test set | L2CD x10^4: 4.221 -> 4.976 | Hurt |
| combine multiple noisy observations by concatenation instead of separate training (ablation) | point cloud denoising and reconstruction · PU test set | L2CD x10^4: denoise 4.221 vs 4.244; reconstruction 4.355 vs 4.315 | Mixed |
| EMD noise-to-noise mapping with geometric consistency vs SBP baseline | point cloud denoising · PU and PC benchmarks, 1-3% noise | L2CD x10^4: e.g. PU 10K 1%: 2.521 -> 1.060 | Helped |
| noise-to-noise SDF learning vs POCO and IMLS on ShapeNet | surface reconstruction · ShapeNet 13 classes | L1CD x10 / NC / F-Score: L1CD 0.030 -> 0.026; NC 0.950 -> 0.962; F-Score 0.984 -> 0.991 (vs POCO) | Helped |
| single noisy point cloud training (N=1) vs NP/IMLS/PCP on ABC/FAMOUS | surface reconstruction · ABC var | L2CD x100: 0.57 (IMLS) -> 0.113 | Helped |
Closest papers by meaning
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (NeurIPS 2024 · similarity 0.91)
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 2025 · similarity 0.88)
- Toward Robust Neural Reconstruction from Sparse Point Sets (CVPR 2025 · similarity 0.88)
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.87)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.87)
- StraightPCF: Straight Point Cloud Filtering (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.