Research map / Point clouds / Shape and surface completion
A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling
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 PUDM, a conditional DDPM for point cloud upsampling with a dual-mapping C-Net/N-Net design and a rate embedding for arbitrary-scale generation, reporting state-of-the-art CD/HD on PUGAN and PU1K.
Problem
Point cloud upsampling relies on sparse-feature encoders and hand-designed upsampling modules that struggle to recover fine geometry, are noise-sensitive due to CD loss, and do not generalize across arbitrary upsampling rates.
Approach
Treat upsampling as conditional generation with a DDPM that models the score of dense point clouds conditioned on the sparse input, using a dual mapping (a conditional branch trained with MSE to reconstruct the sparse input, plus a noise branch) and a rate embedding to handle arbitrary scales, with no explicit upsampling module.
Key results
- On PUGAN at 4x, CD 0.131 and HD 1.220 vs. Grad-PU's 0.245 and 2.369 (CD×10^-3).
- On PU1K at 4x, CD 0.217 vs. Grad-PU's 0.404.
- Under Gaussian noise tau=0.02 at 4x on PUGAN, CD 0.529 vs. Grad-PU's 0.766.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PUDM instead of Grad-PU | PUGAN 4× · PUGAN 4× | CD × 10−3: 0.245 -> 0.131 | Helped |
| Use PUDM instead of Grad-PU | PUGAN 4× · PUGAN 4× | HD × 10−3: 2.369 -> 1.220 | Helped |
| Use PUDM instead of Grad-PU | PUGAN 16× · PUGAN 16× | CD × 10−3: 0.108 -> 0.082 | Helped |
| Use PUDM instead of Grad-PU | PUGAN 16× · PUGAN 16× | HD × 10−3: 2.352 -> 1.120 | Helped |
| Use PUDM instead of Grad-PU | PU1K 4× · PU1K 4× | CD × 10−3: 0.404 -> 0.217 | Helped |
| Use PUDM instead of Grad-PU | PU1K 4× · PU1K 4× | HD × 10−3: 3.732 -> 2.164 | Helped |
| Add dual mapping supervision (ablation) | PUDM · PUGAN 4× | CD × 10−3: 0.168 -> 0.131 | Helped |
| Add dual mapping supervision (ablation) | PUDM · PUGAN 4× | HD × 10−3: 1.498 -> 1.220 | Helped |
| Add the C-Net (ablation) | PUDM · PUGAN 4× | CD × 10−3: 0.212 -> 0.131 | Helped |
| Add rate modeling (ablation) | PUDM · PUGAN, multiple upsampling rates | CD × 10−3: 2×: 0.295 -> 0.247; 4×: 0.158 -> 0.131; 8×: 0.139 -> 0.104 | Helped |
| Use multiple Transfer Modules instead of a single module (ablation) | PUDM · PUGAN 4× | CD × 10−3: 0.131 -> 0.129 | Helped |
| Use multiple Transfer Modules instead of a single module (ablation) | PUDM · PUGAN 4× | HD × 10−3: 1.220 -> 1.235 | Hurt |
Surprising findings
- Performance under uniform noise is notably worse than under Gaussian noise, which the authors attribute to the score-matching objective learning the Gaussian noise direction.
Limitations
- Performs poorly on noise distributions other than Gaussian (e.g., uniform noise).
- P2F is slightly behind Grad-PU at 4x on PUGAN.
- Inference uses 1000 diffusion steps, which implies slow sampling (not quantified in main text).
Experimental setup
Datasets: PUGAN PU1K ScanNet KITTI ModelNet40 ShapeNet
Benchmarks: PUGAN PU1K ScanNet KITTI ModelNet40 ShapeNet
Metrics: Chamfer Distance Hausdorff Distance Point-to-Surface Distance classification instance accuracy classification class accuracy
Baselines: PU-Net MPU PU-GAN Dis-PU PU-EVA PU-GCN NePS Grad-PU
Builds on: DDPM PointNet++ Grad-PU PU-Net
Closest papers by meaning
- Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · similarity 0.91)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.89)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.88)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.87)
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 2025 · similarity 0.87)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 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.