Research map / Point clouds / Shape and surface completion
You Should Learn to Stop Denoising on Point Clouds in Advance
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 ASDN, a U-Net point cloud denoiser with an adaptive classifier that predicts an entropy-based recoverability factor to decide early stopping of denoising per patch.
Problem
Learning-based point cloud denoising tends to over-denoise already-clean points and under-denoise high-noise regions because the network focuses on high-noise points and treats denoising as a homogeneous smoothing process.
Approach
Measure denoising difficulty with a recoverability factor (ratio of voxel-entropy of noisy vs clean point cloud), predict it with an EdgeConv-based adaptive classifier, and map it to a stopping layer in a U-Net so well-denoised patches exit the encoder early while noisy ones go deeper.
Key results
- On PUNet at 10K points with 3% noise, ASDN reaches CD 3.079 vs 4.241 for IterativePFN (x10^-4).
- On Kinect v2, ASDN achieves CD 18.22 and P2M 10.38, better than IterativePFN (18.69 / 10.92).
- Ablation: removing the adaptive classifier raises 3% noise CD on 10K points from 3.079 to 3.580.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ASDN instead of the strongest listed baseline | Point cloud denoising · PUNet 10K, 2% noise | CD and P2M: CD: 3.043 -> 2.581; P2M: 0.555 -> 0.515 | Helped |
| use ASDN instead of the strongest listed baseline | Point cloud denoising · PUNet 10K, 3% noise | CD and P2M: CD: 4.075 -> 3.079; P2M: 1.242 -> 0.946 | Helped |
| use ASDN instead of the strongest listed baseline | Point cloud denoising · Kinect v2 | CD: 18.69 -> 18.22 | Helped |
| use ASDN instead of the strongest listed baseline | Point cloud denoising · Kinect v2 | P2M: 10.29 -> 10.38 | Hurt |
| add the adaptive classifier (ablation) | ASDN · PUNet 10K, 1%, 2%, and 3% noise | CD and P2M: CD: 1.973 -> 1.880; 2.794 -> 2.581; 3.580 -> 3.079; P2M: 0.230 -> 0.219; 0.555 -> 0.515; 1.108 -> 0.946 | Helped |
| increase initial feature dimension from 32 to 64 (ablation) | ASDN · PUNet 10K, 1%, 2%, and 3% noise | CD and P2M: CD: 1.880 -> 1.870; 2.581 -> 2.575; 3.079 -> 2.976; P2M: 0.219 -> 0.222; 0.515 -> 0.526; 0.946 -> 1.023 | Mixed |
| increase the number of layers from 3 to 4 (ablation) | ASDN · PUNet 10K, 1%, 2%, and 3% noise | CD and P2M: CD: 2.016 -> 1.880; 2.785 -> 2.581; 3.829 -> 3.079; P2M: 0.245 -> 0.219; 0.556 -> 0.515; 1.317 -> 0.946 | Helped |
| increase the number of layers from 4 to 5 (ablation) | ASDN · PUNet 10K, 1%, 2%, and 3% noise | CD and P2M: CD: 1.880 -> 1.865; 2.581 -> 2.532; 3.079 -> 2.936; P2M: 0.219 -> 0.247; 0.515 -> 0.549; 0.946 -> 1.134 | Mixed |
Limitations
- Recoverability factor relies on a ground-truth clean point cloud entropy during training, and the theoretical foundation of entropy is not analyzed.
- Entropy depends on voxel size choice and the mapping parameter gamma is hand-set.
- Evaluated mainly on PUNet with Gaussian noise; real-scan evaluation is qualitative on one street scene.
Experimental setup
Datasets: PUNet Kinect v2 dataset Paris-rue-Madame
Benchmarks: PUNet Kinect v2 Paris-rue-Madame
Metrics: CD P2M
Baselines: PCN GPDNet DMR PointFilter Score PDFlow DeepPSR MAG IterativePFN
Builds on: IterativePFN DGCNN PointCleanNet InfoCD loss
Closest papers by meaning
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 2025 · similarity 0.91)
- Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · similarity 0.89)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.87)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.86)
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · similarity 0.86)
- Denoising Point Clouds in Latent Space via Graph Convolution and Invertible Neural Network (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.