Research map / Point clouds / Shape and surface completion
SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh 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 SPU-PMD, a self-supervised point cloud upsampling network that combines geometric mesh interpolation with progressive mesh deformation (recurrent feature aggregation and gated coordinate reconstruction) to densify point clouds.
Problem
Point cloud upsampling: generating dense, uniformly distributed, structurally detailed point clouds from sparse inputs, especially without paired dense ground truth (e.g., real scans).
Approach
Treats upsampling as inferring dense points from a deformable mesh: an interpolation step inserts mesh centroids, then cascaded deformers (RFA recurrent memory + GCR gated motion) move mesh nodes to refine uniformity and structure, trained self-supervised with Chamfer and uniformity losses against the sparse input.
Key results
- On PU1K (dense 2,048 input), SPU-PMD achieves CD 0.544e-3, HD 4.926e-3, P2F 1.861e-3 vs PU-CRN's CD 0.499e-3 and HD 8.068e-3, so it leads in HD and P2F but not CD.
- Trained on PU1K and tested on PU-GAN, SPU-PMD gives the best HD and P2F at all densities, with CD only behind the supervised PU-CRN.
- Ablation on PU1K shows that replacing mesh interpolation with linear interpolation or point duplication degrades CD and HD substantially (e.g., LI dense CD 12.994e-3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPU-PMD instead of PU-CRN | Point cloud upsampling · PU1K sparse (512) | HD (10^-3): 18.835 -> 13.568 | Helped |
| Use SPU-PMD instead of PU-CRN | Point cloud upsampling · PU1K sparse (512) | CD (10^-3): 1.611 -> 1.690 | Hurt |
| Remove RFA (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 1.722 | Hurt |
| Replace RFA with GRU (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 1.711 | Hurt |
| Remove GCR and retain coordinate regression (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 1.716 | Hurt |
| Replace mesh interpolation with linear interpolation (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 11.228 | Hurt |
| Replace mesh interpolation with point duplication (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 3.162 | Hurt |
| Add uniform loss to the loss configuration (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 2.000 -> 1.965 | Helped |
| Apply Lcd in each stage instead of the baseline loss arrangement (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 2.000 | Hurt |
| Apply Lcd + Lu in each stage instead of the baseline loss arrangement (ablation) | SPU-PMD · PU1K sparse (512) | CD (10^-3): 1.690 -> 1.965 | Hurt |
| Add mesh deformation stages (ablation) | SPU-PMD · PU1K point clouds | CD (10^-3): 1.586 -> 0.544 | Helped |
| Add mesh deformation stages (ablation) | SPU-PMD · PU1K point clouds | P2F avg (10^-3): 1.503 -> 1.861 | Hurt |
Surprising findings
- Linear interpolation and point duplication yield the lowest P2F values, which the authors attribute to points lying on the same planes, though their CD and HD are much worse.
Limitations
- Supervised baseline PU-CRN still achieves better CD on PU1K and PU-GAN.
- Interpolation-based initialization yields coarse, non-uniform points that depend on the deformers to correct them.
Experimental setup
Datasets: PU1K PU-GAN
Benchmarks: PU1K PU-GAN KITTI
Metrics: Chamfer distance Hausdorff distance point-to-surface distance
Baselines: PU-Net MPU PU-GAN Dis-PU PU-GCN PU-CRN SAPCU SPU-Net
Builds on: PU-Net Point Transformer SPU-Net GRU
Closest papers by meaning
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.90)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.89)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.87)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.87)
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.87)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 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.