Research map / Point clouds / Shape and surface completion
NoiseSDF2NoiseSDF: Learning Clean Neural Fields from Noisy Supervision
Tengkai Wang, Weihao Li, Ruikai Cui, Shi Qiu, Nick Barnes
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train with noisy SDF targets (NoiseSDF2NoiseSDF) on top of 3DS2V | 3DS2V · ShapeNet (sigma=0.01) | IoU: 0.884 -> 0.929 | Improved |
| train with noisy SDF targets on top of 3DS2V at higher noise | 3DS2V · ShapeNet (sigma=0.02) | Normal Consistency: 0.733 -> 0.868 | Improved |
| train with noisy SDF targets under Laplace noise | 3DS2V · ShapeNet Chair Laplace (sigma=0.01) | IoU: 0.909 -> 0.908 | No clear effect |
| train with noisy SDF targets under Gaussian noise with mean 0.02 | 3DS2V · ShapeNet Chair Gaussian (mu=0.02) | F-Score: 0.549 -> 0.524 | Hurt |
| use noisy targets in place of clean SDF targets (ablation) | NoiseSDF2NoiseSDF · ShapeNet Chair (3DS2V network) | IoU: 0.939 (clean) vs 0.927 (noisy) | No clear effect |
| use 3DILG instead of 3DS2V to generate noisy targets (ablation) | NoiseSDF2NoiseSDF · ShapeNet Chair (3DS2V network) | IoU: 0.927 -> 0.903 | Hurt |
| add cross-attention block to fine-tuned FC layer (ablation) | denoising network · ShapeNet Chair | IoU: 0.884 -> 0.922 | Improved |
| fine-tune the whole decoder instead of FC+CA (ablation) | denoising network · ShapeNet Chair | IoU: 0.922 -> 0.927 | Improved |
| randomly initialize the whole network instead of 3DS2V init (ablation) | denoising network · ShapeNet Chair | IoU: 0.927 -> 0.929 | No clear effect |
| average three noisy targets (1-to-3) instead of one (ablation) | NoiseSDF2NoiseSDF · ShapeNet Chair | IoU: 0.927 -> 0.923 | No clear effect |
| add test-time optimization with 200 noisy samples | NoiseSDF2NoiseSDF · Famous medium noise | Chamfer: 0.0160 -> 0.0108 | Improved |
| use end-to-end noisy-SDF training instead of denoise-then-reconstruct | NoiseSDF2NoiseSDF · ShapeNet Chair | IoU: 0.882 -> 0.927 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- POCO — POCO: Point Convolution for Surface Reconstruction (CVPR 2022) · better on 0 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Neural Kernel Surface Reconstruction (CVPR 2023 · similarity 0.89)
- Fast Learning of Signed Distance Functions From Noisy Point Clouds via Noise to Noise Mapping (TPAMI 2024 · similarity 0.89)
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 2025 · similarity 0.88)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.88)
- NeuralIndicator: Implicit Surface Reconstruction from Neural Indicator Priors (ICML 2024 · similarity 0.87)
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (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.
