Research map / Point clouds / Shape and surface completion
StraightPCF: Straight Point Cloud Filtering
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 StraightPCF, a lightweight (~530K params) point cloud filtering network that moves noisy points along straight flow trajectories via coupled VelocityModules and a DistanceModule scalar, achieving state-of-the-art CD/P2M on synthetic and real scanned data.
Problem
Deep point cloud filtering methods move noisy points along stochastic trajectories, requiring regularization and many iterations, and produce clustered point distributions and large networks.
Approach
Model noisy patches as intermediate states of a linear interpolation between high-noise and clean patches, learn a constant flow velocity (straight path) with a DGCNN-based VelocityModule, straighten flows by coupling K=2 VelocityModules, and scale the step with a learned DistanceModule scalar so points converge to the surface.
Key results
- On PCNet at 10K resolution and 3% noise, CD is reduced by 17.1% and P2M by 24.1% relative to the next best method.
- Achieves ~530K parameters, about 17% of IterativePFN's 3.2M parameters.
- On Kinect data, CD of 1.296 vs 1.320 for IterativePFN (1.82% reduction).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use coupled VelocityModules with a DistanceModule rather than a single VelocityModule without it (ablation) | StraightPCF · PUNet, 10K, 3% noise | CD: 3.73 -> 3.29 | Helped |
| Add the DistanceModule to a single VelocityModule (ablation) | StraightPCF · PUNet, 10K, 3% noise | CD: 3.06 -> 3.81 | Hurt |
| Increase the size of the single VelocityModule (ablation) | StraightPCF · PUNet, 10K, 3% noise | CD: 3.73 -> 3.54 | Helped |
| Use coupled VelocityModules without the DistanceModule instead of a single VelocityModule without it (ablation) | StraightPCF · PUNet, 10K, 3% noise | CD: 3.73 -> 3.71 | No clear effect |
| Use coupled VelocityModules with a DistanceModule rather than coupled VelocityModules without it (ablation) | StraightPCF · PUNet, 10K, 3% noise | CD: 3.71 -> 3.29 | Helped |
| Use StraightPCF instead of DeepPSR | Point cloud filtering · PUNet, 10K, 3% noise | CD: 4.075 -> 3.287 | Helped |
| Use StraightPCF instead of DeepPSR | Point cloud filtering · PUNet, 50K, 3% noise | CD: 1.344 -> 1.307 | Helped |
| Use StraightPCF instead of IterativePFN | Point cloud filtering · PCNet, 10K, 3% noise | CD: 6.026 -> 4.921 | Helped |
| Use StraightPCF instead of DeepPSR | Point cloud filtering · PCNet, 50K, 3% noise | CD: 2.093 -> 1.816 | Helped |
| Use StraightPCF instead of IterativePFN | Point cloud filtering · Kinect | CD: 1.320 -> 1.296 | Helped |
Surprising findings
- Simply increasing parameter count of a single VelocityModule (Large VM) yields only limited gains compared with the coupled VM + DM design.
Limitations
- Relatively low performance on low-density or highly sparse point clouds, similar to current methods.
- Filtering starting time must be implicitly modeled; too many Euler steps may cause non-convergence.
Experimental setup
Datasets: PUNet dataset PCNet dataset Kinect v1 dataset Paris-Rue-Madame dataset
Benchmarks: PUNet dataset PCNet dataset Kinect v1 dataset Paris-Rue-Madame dataset
Metrics: Chamfer Distance (CD) Point2Mesh (P2M)
Baselines: PCN PointFilter ScoreDenoise PDFlow DeepPSR IterativePFN
Builds on: ScoreDenoise IterativePFN Reflow DGCNN
Closest papers by meaning
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.89)
- SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors (NeurIPS 2024 · similarity 0.85)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.85)
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 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.