Research map / Point clouds / Shape and surface completion
Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning
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 PU-VoxelNet, a voxel-based arbitrary-scale point cloud upsampling network with density-guided grid resampling (D-FPS) and a latent geometric-consistent auxiliary loss using a pretrained EdgeConv surface encoder.
Problem
Arbitrary-scale point cloud upsampling from sparse inputs; point-based surface approximation yields low-fidelity geometry, and existing voxel grid sampling suffers from inaccurate sampling with outliers at large rates; CD loss ignores local point distribution.
Approach
Model the local surface as a per-cell point density on a predefined voxel grid, resample grid cells via multinomial sampling followed by density-weighted farthest point sampling (D-FPS), and add a latent-space loss that compares EdgeConv surface codes of real vs. seed-replaced local patches.
Key results
- On PU1K 4x upsampling, CD 0.338, HD 2.694, P2F 1.183 (x10^-3), beating Grad-PU (CD 0.404).
- On PU-GAN 16x upsampling, CD 0.091 vs. Grad-PU 0.108 (x10^-3).
- Surface reconstruction CD on PU-GAN 0.245 (x10^-3), best among compared methods except High-res reference (0.203).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PU-VoxelNet instead of Grad-PU | Point cloud upsampling · PU1K, 4× upsampling | CD (×10³): 0.404 -> 0.338 | Helped |
| Use PU-VoxelNet instead of Grad-PU | Point cloud upsampling · PU-GAN, 4× upsampling | P2F (×10³): 1.893 -> 2.137 | Hurt |
| Use PU-VoxelNet instead of Grad-PU | Point cloud upsampling · PU-GAN, 16× upsampling | P2F (×10³): 2.127 -> 2.301 | Hurt |
| Use density-guided FPS instead of vanilla FPS (ablation) | PU-VoxelNet · PU1K, 4× upsampling | CD (×10³): 0.366 -> 0.352 | Helped |
| Use density-guided FPS instead of vanilla FPS (ablation) | PU-VoxelNet · PU1K, 16× upsampling | CD (×10³): 0.433 -> 0.306 | Helped |
| Add latent geometric-consistent learning (ablation) | PU-VoxelNet · PU1K, Poisson sampling | CD (×10³): 0.352 -> 0.338 | Helped |
| Add latent geometric-consistent learning (ablation) | PU-VoxelNet · PU1K, random sampling | CD (×10³): 0.506 -> 0.475 | Helped |
| Replace latent geometric-consistent learning with SurRep (ablation) | PU-VoxelNet · PU1K, Poisson sampling | HD (×10³): 2.694 vs 2.646 | Mixed |
| Remove sharp CD loss (ablation) | PU-VoxelNet · PU1K | CD (×10³): 0.338 -> 0.349 | Hurt |
| Remove regularization loss (ablation) | PU-VoxelNet · PU1K | CD (×10³): 0.338 -> 0.351 | Hurt |
| Remove occupancy BCE loss (ablation) | PU-VoxelNet · PU1K | CD (×10³): 0.338 -> 0.408 | Hurt |
| Remove density MSE loss (ablation) | PU-VoxelNet · PU1K | CD (×10³): 0.338 -> 0.389 | Hurt |
Surprising findings
- Voxel-based generation baselines (CDA, GRNet, VE-PCN) perform poorly on upsampling, indicating generic voxel generators are unsuitable for this task.
- Own method's PU1K 4x CD/HD/P2F in the main table is compared against 2,048-point inputs; ablation shows the GC loss also helps MPU and PU-GCN.
Limitations
- Relies on imprecise density and occupancy predictions that can cause inaccurate sampling.
- Latent geometric loss depends on a pretrained surface encoder and ground-truth neighbors during training.
Experimental setup
Datasets: PU1K PU-GAN ScanObjectNN KITTI
Benchmarks: PU1K PU-GAN ScanObjectNN KITTI
Metrics: Chamfer Distance Hausdorff Distance Point-to-Surface Distance
Baselines: PU-Net MPU PU-GAN Dis-PU PU-GCN MAFU PU-SSAS Grad-PU CDA GRNet VE-PCN
Builds on: PU-GCN Grad-PU PU-GAN EdgeConv Chamfer Distance
Closest papers by meaning
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · similarity 0.89)
- 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 0.89)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.88)
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.87)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (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.