Research map / Point clouds / Shape and surface completion
SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network
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.
Frames point cloud upsampling as masked shape completion: an iterative mask-recovery transformer, trained self-supervised with EMD, fills masked patches and merges multiple mask-sequence outputs for arbitrary-scale upsampling.
Problem
Point cloud upsampling (generating dense, uniform point sets from sparse inputs) is usually cast as local interpolation, which struggles with local topology and typically needs dense supervision; the paper targets self-supervised, arbitrary-scale upsampling.
Approach
Treat upsampling as global shape completion: mask point-cloud patches, iteratively deform learnable patch embeddings guided by visible-patch features and center-point positional encodings, and at test time combine outputs from multiple random mask sequences (MMR module) followed by outlier removal and FPS to reach any ratio.
Key results
- At 2x, the method achieves the best EMD (24.3), HD (25.2) and F-Score (49.1) among all compared methods in the table, per the reported ranking.
- At 8x, the method attains the best F-Score (49.2) and P2F (5.31) and the best overall ranking on several metrics.
- Ablation on mask ratio shows m=0.6 gives the best trade-off (CD 13.9, EMD 22.2, HD 22.7 at 4x); performance saturates at L=3 iteration layers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPU-IMR instead of PUSS-AS | arbitrary-scale upsampling · Upsampling 2× | HD: 32.2 -> 25.2 | Helped |
| Use SPU-IMR instead of Grad-PU | arbitrary-scale upsampling · Upsampling 4× | P2F: 7.19 -> 6.85 | Helped |
| Use SPU-IMR instead of Grad-PU | arbitrary-scale upsampling · Upsampling 4× | F-Score: 49.5 -> 49.2 | Hurt |
| Use SPU-IMR instead of PUSS-AS | arbitrary-scale upsampling · Upsampling 8× | P2F: 5.40 -> 5.31 | Helped |
| Use mask ratio 0.3 instead of 0.6 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.7 | Hurt |
| Use mask ratio 0.4 instead of 0.6 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 15.3 | Hurt |
| Use mask ratio 0.5 instead of 0.6 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.1 | Hurt |
| Use mask ratio 0.7 instead of 0.6 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.5 | Hurt |
| Use mask ratio 0.8 instead of 0.6 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.3 | Hurt |
| Use 1 iteration instead of 3 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 15.9 | Hurt |
| Use 2 iterations instead of 3 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.0 | No clear effect |
| Use 4 iterations instead of 3 (ablation) | SPU-IMR · Upsampling 4× | CD: 13.9 -> 14.5 | Hurt |
Surprising findings
- Very low mask ratios (0.3, 0.4) cause clustering of generated points and abnormal metrics, attributed to overfitting from abundant known information.
Limitations
- Self-supervised model is not uniformly best; CD and some metrics trail fully supervised methods such as PU-GCN and Grad-PU at certain ratios.
- Test-time MMR requires multiple mask sequences (q=8) and outlier removal plus FPS, adding inference cost.
Experimental setup
Datasets: PU-GAN dataset (4719 point clouds, 3672 train / 1047 test)
Benchmarks: PU-GAN test set PU1K KITTI
Metrics: CD EMD HD F-Score P2F NUC
Baselines: PU-Net MPU PU-GAN PU-GCN Grad-PU SAPCU PUSS-AS
Builds on: Point-MAE PointTr PU-Net SAPCU
Closest papers by meaning
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.91)
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · 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 0.88)
- RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation (CVPR 2024 · similarity 0.88)
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · similarity 0.88)
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.