Research map / Point clouds / Shape and surface completion
Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression
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 PDANS, a conditional diffusion network for point cloud upsampling with an Adaptive Noise Suppression module and a TreeTrans decoder, reporting gains over PUDM and other baselines on PUGAN and PU1K including noisy inputs.
Problem
Point cloud upsampling degrades under noisy inputs because existing methods rely on outlier-sensitive losses (CD/EMD) and FPS-based sampling that retains noisy points.
Approach
Combines a conditional DDPM (no CD/EMD loss) with an ANS module that computes per-point attention weights over KNN neighborhoods to remove noisy points, plus a TreeTrans decoder with ancestral/recurrent terms and local/global attention, and a cross-attention link between a sparse-point conditional network and the diffusion network.
Key results
- On PUGAN at 4x, CD 0.119 vs 0.131 for PUDM (best prior); at 16x CD 0.078 vs 0.082 for PUDM.
- On PU1K at 4x, CD 0.155 vs 0.217 for PUDM, HD 1.588 vs 2.164.
- Under Gaussian noise tau=0.05 at 4x on PUGAN, CD 0.476 vs 0.618 for PUDM; ablation shows removing TT raises 16x CD from 0.078 to 0.094.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PDANS instead of PUDM | Point cloud upsampling · PUGAN 4× | CD ×10−3: 0.131 -> 0.119 | Helped |
| use PDANS instead of PUDM | Point cloud upsampling · PUGAN 16× | CD ×10−3: 0.082 -> 0.078 | Helped |
| use PDANS instead of PUDM | Point cloud upsampling · PU1K 4× | CD ×10−3: 0.217 -> 0.155 | Helped |
| use PDANS instead of PUDM | Point cloud upsampling · PUGAN Gaussian-noise tests, τ=0.01–0.1 | CD ×10−3: 0.210 -> 0.185; 0.529 -> 0.309; 0.618 -> 0.476; 0.853 -> 0.543 | Helped |
| use PDANS instead of PUDM | Point cloud upsampling · PUGAN random-noise tests, τ=0.05 and τ=0.1 | CD ×10−3: 0.998 -> 0.722; 1.310 -> 0.682 | Helped |
| use PDANS instead of PUDM | Point cloud upsampling · PUGAN Laplace-noise tests, τ=0.05 and τ=0.1 | CD ×10−3: 0.643 -> 0.533; 0.704 -> 0.563 | Helped |
| replace ANS with Dense Blocks (ablation) | PDANS encoder · PUGAN 4× Gaussian noise τ=0.1 | CD ×10−3: 0.543 -> 1.248 | Hurt |
| replace ANS with Set Abstraction (ablation) | PDANS encoder · PUGAN 4× Gaussian noise τ=0.1 | CD ×10−3: 0.543 -> 0.853 | Hurt |
| replace ANS with PointNet++ encoder (ablation) | PDANS encoder · PUGAN 4× Gaussian noise τ=0.1 | CD ×10−3: 0.543 -> 1.168 | Hurt |
| remove TreeTrans module (ablation) | PDANS · PUGAN 16× | CD ×10−3: 0.078 -> 0.094 | Hurt |
Surprising findings
- On PUGAN at 4x, the proposed method has a slightly higher HD (1.2065) than PUDM (1.220 is higher, so not) — none clearly reported; P2F is not best in several settings (e.g., PU1K 4x P2F 1.354 vs RepKPU 1.261).
Limitations
- Performance under random and Laplacian noise is noticeably lower than under Gaussian noise.
- P2F is not consistently best across settings.
- Evaluation relies on synthetic noise injection with only limited real-scan qualitative comparison (no ground truth on ScanObjectNN).
Experimental setup
Datasets: PUGAN PU1K
Benchmarks: PUGAN PU1K ScanObjectNN
Metrics: CD HD P2F
Baselines: PU-Net MPU PU-GAN PU-GCN Grad-PU RepKPU PUDM
Builds on: DDPM DDIM PUDM Tree-GCN
Closest papers by meaning
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · similarity 0.91)
- You Should Learn to Stop Denoising on Point Clouds in Advance (AAAI 2025 · similarity 0.89)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.88)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.86)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.86)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · 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.