Research map / Point clouds / Shape and surface completion
RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation
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 RepKPU, a point cloud upsampling network that encodes local geometry as kernel points (RepKPoints) and generates points by deforming KP-Queries via cross-attention (Kernel-to-Displacement).
Problem
Point cloud upsampling suffers from outliers, shrinkage, and limited local shape representations when generating dense points from sparse inputs.
Approach
Represents local geometry with deformable KPConv-style kernel points (RepKPoints) and reformulates upsampling as deforming a second set of predefined kernel-point queries (KP-Queries) through cross-attention to predict displacements.
Key results
- On PU1K 4× upsampling, RepKPU reports CD 0.327e-3, HD 0.268e-2, P2F 0.938e-3, better than Grad-PU (CD 0.404e-3).
- On PU-GAN 16× upsampling, RepKPU achieves CD 0.107e-3 versus Grad-PU's 0.129e-3, and is about 29% faster than the second-fastest MPU at inference.
- Robustness test averages show RepKPU with CD 0.732e-3 overall, outperforming baselines across uniform, noisy, and random inputs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use deformable rather than rigid RepKPoints (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.345 -> 0.327 | Helped |
| Use deformable rather than rigid RepKPoints (ablation) | RepKPU · PU1K 4× | HD (×10^2): 0.286 -> 0.268 | Helped |
| Use deformable rather than rigid RepKPoints (ablation) | RepKPU · PU1K 4× | P2F (×10^3): 1.118 -> 0.938 | Helped |
| Add REM-extracted Fg to the baseline folding design (ablation) | RepKPU baseline · PU1K 4× | CD (×10^3): 0.365 -> 0.346 | Helped |
| Use local patches instead of RepKPoints (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.354 | Hurt |
| Use local patches instead of RepKPoints and add Fg (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.344 | Hurt |
| Use folding queries instead of proposed KP-Queries generation (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.355 | Hurt |
| Replace rigid folding seeds with a learnable query matrix (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.348 | Hurt |
| Use Graph Convolution instead of Vector Attention in the encoder (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.334 | Hurt |
| Use Set Abstraction instead of Vector Attention in the encoder (ablation) | RepKPU · PU1K 4× | CD (×10^3): 0.327 -> 0.347 | Hurt |
| Apply Fixed-to-Arbitrary (F2A) to RepKPU | RepKPU · PU-GAN 16× | HD (×10^3): 3.345 -> 1.975 | Helped |
Surprising findings
- Grad-PU fails on random inputs in the robustness test, with average CD worse than several earlier methods.
Limitations
- Arbitrary-scale upsampling relies on a heuristic F2A strategy with midpoint interpolation and random point dropping rather than a learned approach.
- Fixed-scale training at 4x; 16x is obtained by repeated 4x or F2A.
Experimental setup
Datasets: PU-GAN dataset PU1K dataset
Benchmarks: PU-GAN PU1K ScanObjectNN
Metrics: Chamfer Distance Hausdorff distance Point-to-surface Distance seconds per point cloud training memory
Baselines: PU-Net MPU PU-GAN Dis-PU PU-GCN Grad-PU PU-Flow MAFU
Builds on: KPConv Deformable KPConv Vector attention Grad-PU PU-GCN
Closest papers by meaning
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.88)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.88)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.86)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.86)
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · similarity 0.85)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 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.