Research map / Point clouds / Shape and surface completion
PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale 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 PLATYPUS, which uses a Progressive Local Surface Estimator with curvature-based sampling and a curvature-distribution-based curriculum learning strategy to learn an unsigned distance field for arbitrary-scale point cloud upsampling.
Problem
Point cloud upsampling via projecting query points onto the underlying surface struggles with outliers, non-uniformity, and high-curvature or intricate regions.
Approach
Curvature-based progressive sampling (top-k highest-curvature points at k/2, k/4, k/8, k/16) feeds multi-scale features into a distance estimator, and the global mean curvature of each training sample defines easy/hard curriculum stages for the first time on point clouds.
Key results
- On PU-GAN 4x, PLATYPUS achieves CD 0.229 and HD 1.426 (x10^-3), best among compared methods on CD and HD.
- On PU-GAN 16x, CD 0.088 and HD 1.429 are best; P2F is worse than APU-LDI (2.142 vs 1.544).
- Ablation on PU-GAN 4x: removing both curvature-based sampling and curriculum learning raises CD from 0.229 to 0.258 and HD from 1.426 to 2.720.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 4× | CD: 0.232 -> 0.229 | Helped |
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 4× | HD: 1.675 -> 1.426 | Helped |
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 4× | P2F: 1.338 -> 1.908 | Hurt |
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 16× | CD: 0.092 -> 0.088 | Helped |
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 16× | HD: 1.504 -> 1.429 | Helped |
| Use PLATYPUS instead of APU-LDI | Point cloud upsampling · PU-GAN 16× | P2F: 1.544 -> 2.142 | Hurt |
| Add curvature-based sampling and curriculum learning (ablation) | PLATYPUS · PU-GAN | CD: 0.258 -> 0.229 | Helped |
| Add curvature-based sampling and curriculum learning (ablation) | PLATYPUS · PU-GAN | HD: 2.720 -> 1.426 | Helped |
| Add curvature-based sampling and curriculum learning (ablation) | PLATYPUS · PU-GAN | P2F: 2.013 -> 1.908 | Helped |
| Add curvature-based sampling (ablation) | PLATYPUS without curriculum learning · PU-GAN | CD: 0.258 -> 0.235 | Helped |
| Add curvature-based sampling (ablation) | PLATYPUS without curriculum learning · PU-GAN | HD: 2.720 -> 1.739 | Helped |
| Add curvature-based sampling (ablation) | PLATYPUS without curriculum learning · PU-GAN | P2F: 2.013 -> 1.983 | Helped |
Surprising findings
- Under Gaussian noise with tau=0.01, PLATYPUS has a worse CD (0.368) and P2F (6.804) than APU-LDI and some baselines, despite the best HD (2.224).
Limitations
- Authors note room for optimizing memory usage for storing sampled point clouds.
- P2F is not best on several settings (e.g., 4x P2F 1.908 vs APU-LDI 1.338).
Experimental setup
Datasets: PU-GAN PU1K
Benchmarks: PU-GAN PU1K KITTI ScanObjectNN
Metrics: Chamfer distance Hausdorff distance point-to-surface distance
Baselines: PU-Net MPU PU-GAN Dis-PU PU-GCN NePs Grad-PU APU-LDI PU-VoxelNet RepKPU MAFU PC2-PU
Builds on: Grad-PU PU-GAN APU-LDI
Closest papers by meaning
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.90)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.88)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.88)
- CurveCloudNet: Processing Point Clouds with 1D Structure (CVPR 2024 · similarity 0.87)
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · similarity 0.87)
- MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing (CVPR 2025 · 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.