Research map / Point clouds / Shape and surface completion
Learning Continuous Implicit Field with Local Distance Indicator 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.
Learns a global unsigned distance field for point cloud upsampling, guided by an attention-based local distance indicator trained on dense local patches, enabling arbitrary-scale upsampling by projecting random queries onto the zero level set.
Problem
Patch-based point cloud upsampling produces holes, outliers and non-uniformity because independently upsampled local patches are inconsistent when merged; arbitrary-scale upsampling also typically needs retraining.
Approach
Train a local distance indicator (LDI) with attention over neighboring patch points to predict query-to-surface distance, then use it as a prior to supervise a global unsigned distance field over the whole sparse cloud; upsampling projects random queries onto the implicit surface, and the continuous field allows any scale without retraining.
Key results
- At 4x on PU-GAN, CD 0.232e-3, HD 1.675e-3, P2F 1.338e-3, beating Grad-PU and PC2-PU on all metrics.
- At 16x on PU-GAN, CD 0.092e-3 and HD 1.504e-3, improving over PC2-PU (CD 0.111) and Grad-PU (HD 2.421).
- On PU1K at 4x, CD 0.371e-3 vs. 0.404e-3 for the next best (Grad-PU).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the continuous implicit field with the proposed LDI | Point cloud upsampling · PU-GAN 4× | CD (10^-3): 0.260 -> 0.232 | Helped |
| Use the continuous implicit field with the proposed LDI | Point cloud upsampling · PU1K 4× | CD (10^-3): 0.404 -> 0.371 | Helped |
| Replace inverse-distance interpolation with attention weighting (ablation) | Local distance indicator with gradient descent · PU1K subset | CD (10^-3): 0.759 -> 0.678 | Helped |
| Replace inverse-distance interpolation with attention weighting (ablation) | Global implicit field · PU1K subset | CD (10^-3): 0.714 -> 0.637 | Helped |
| Use the global implicit field instead of gradient descent (ablation) | Attention-based indicator · PU1K subset | CD (10^-3): 0.678 -> 0.637 | Helped |
| Use the global implicit field instead of gradient descent (ablation) | Interpolation-based indicator · PU1K subset | CD (10^-3): 0.759 -> 0.714 | Helped |
| Use the proposed design instead of OnSurfacePrior (ablation) | Point cloud upsampling · PU1K subset | HD (10^-3): 9.533 -> 4.864 | Helped |
| Add the local distance loss (ablation) | Full method · PU1K subset | CD (10^-3): 0.874 -> 0.637 | Helped |
| Add the nearest-point loss (ablation) | Full method · PU1K subset | CD (10^-3): 0.637 -> 0.555 | Hurt |
| Add the nearest-point loss (ablation) | Full method · PU1K subset | HD (10^-3): 5.720 -> 4.864 | Helped |
| Add the shortest-path loss (ablation) | Full method · PU1K subset | CD (10^-3): 0.746 -> 0.637 | Helped |
| Add the shortest-path loss (ablation) | Full method · PU1K subset | P2F (10^-3): 1.344 -> 1.401 | Hurt |
Surprising findings
- Naively transferring OnSurfacePrior to upsampling gives much worse HD (9.533e-3), indicating the proposed design matters.
- The nearest-point loss worsens CD in the ablation (w/o L_np gives lowest CD of 0.555e-3) despite improving other metrics.
Limitations
- Ablation study uses a subset of PU1K, and some baselines were retrained with default settings rather than official weights for a fair comparison.
- Real-scan evaluation is qualitative only because ground truth is unavailable.
Experimental setup
Datasets: PU-GAN dataset PU1K dataset
Benchmarks: PU-GAN dataset PU1K dataset ScanObjectNN (qualitative) KITTI (qualitative)
Metrics: CD HD P2F
Baselines: PU-Net MPU PU-GAN Dis-PU PU-GCN NePs PU-SMG PC2-PU MAFU Grad-PU PU-Transformer OnSurfacePrior
Builds on: Grad-PU OnSurfacePrior DeepSDF PU-Net
Closest papers by meaning
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.91)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.90)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.89)
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · similarity 0.89)
- RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation (CVPR 2024 · similarity 0.88)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · 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.