Research map / Point clouds / Shape and surface completion
U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching
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 U-CAN, an unsupervised point cloud denoiser trained with EMD-based noise-to-noise matching and a denoising-consistency loss, also transferred to image denoising via ZS-N2N.
Problem
Point cloud denoising usually requires paired noisy-clean data, which is costly to obtain; existing unsupervised methods lack local-level constraints and yield imprecise local geometry.
Approach
Adapts Noise2Noise to unordered point clouds by using Earth Mover's Distance as a one-to-one correspondence-inducing loss between denoised outputs and other noisy observations, plus a consistency loss between denoised predictions of two noisy observations; the denoiser is a multi-step DGCNN-based path predictor.
Key results
- On PUNet 10K with 3% noise, U-CAN reaches CD 3.666 (x10^4), better than the unsupervised TTD, DMR-TTD and ScoreDenoise-TTD baselines.
- On McMaster18 Poisson denoising at lambda=50, U-CAN achieves 31.03 dB PSNR, versus 30.36 dB for ZS-N2N, roughly 1 dB improvement.
- Ablation: removing the consistency loss raises CD from 3.66 to 7.22 at 10K, 3% noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add U-CAN denoising to the unsupervised point-cloud denoising method | PUNet 10K point clouds · PUNet 10K, 2% noise | CD ×10^4: 3.686 -> 3.234 | Helped |
| Add U-CAN denoising to the unsupervised point-cloud denoising method | PUNet 10K point clouds · PUNet 10K, 3% noise | CD ×10^4: 4.708 -> 3.666 | Helped |
| Add the denoising consistency constraint to ZS-N2N | ZS-N2N · McMaster18, Poisson λ=50 | PSNR (dB): 30.36 -> 31.03 | Helped |
| Add the denoising consistency constraint to ZS-N2N | ZS-N2N · McMaster18, Poisson λ=25 | PSNR (dB): 28.41 -> 29.14 | Helped |
| Add the denoising consistency constraint to ZS-N2N | ZS-N2N · McMaster18, Poisson λ=10 | PSNR (dB): 25.75 -> 26.52 | Helped |
| Remove the denoising consistency loss (ablation) | U-CAN · PUNet 10K, 3% noise | CD ×10^4: 3.66 -> 7.22 | Hurt |
| Replace EMD with CD in the noise-to-noise matching loss (ablation) | U-CAN · PUNet point-cloud denoising | denoising performance | Hurt |
| Replace EMD with DCD in the noise-to-noise matching loss (ablation) | U-CAN · PUNet point-cloud denoising | training stability | Hurt |
| Vary the number of denoising steps from 1 to 7 (ablation) | U-CAN · PUNet point-cloud denoising | denoising performance | Mixed |
Surprising findings
- Using Chamfer distance instead of EMD in the noise-to-noise loss leads to suboptimal results and density-aware Chamfer causes severe divergence.
Limitations
- Scanned-data evaluation is qualitative only since no ground truth is available.
- Performance is weaker than supervised methods on some settings, e.g. low-noise 10K cases where TTD and ScoreDenoise-TTD have lower P2M.
Experimental setup
Datasets: PUNet Paris-rue-Madame McMaster18
Benchmarks: PUNet Paris-rue-Madame McMaster18
Metrics: Chamfer distance point-to-mesh distance PSNR
Baselines: Bilateral Jet MRPCA GLR PCNet GPDNet DMR ScoreDenoise TTD DMR-TTD ScoreDenoise-TTD PU-Net
Builds on: Noise2Noise DGCNN ZS-N2N TotalDenoising
Closest papers by meaning
- You Should Learn to Stop Denoising on Point Clouds in Advance (AAAI 2025 · similarity 0.91)
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 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 0.86)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.86)
- Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · similarity 0.86)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.85)
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.