Research map / Point clouds / Point cloud classification
Point Cloud Dataset Distillation
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 DD3D, the first dataset distillation framework for point clouds, combining a sign-invariant PCA-style rotator with a SIREN-based point-wise generator that trains on low-resolution clouds and generates high-resolution ones.
Problem
Dataset distillation has been developed for structured images and graphs, but unstructured point clouds suffer from random orientations (which degrade distillation via weakened principal components) and variable resolutions (fixed-size synthetic tensors are memory-heavy).
Approach
Align point clouds to a canonical orientation with a learned sign-invariant rotator (PCA plus learned reflection), and parameterize synthetic data as a class-conditioned implicit neural representation (SIREN generator) mapping uniform noise to 3D coordinates, so resolution is set by the number of sampled noises.
Key results
- On ModelNet40 (aligned, CPC=50) DD3D reaches 83.91% vs 83.02% for DM and 81.74% for GM, against 88.05% full-dataset accuracy.
- On ShapeNet part segmentation with CPC=1, DD3D reaches 73.06% OA vs 65.56% for GM, and about 90% of full-data performance using 1% of the data.
- Rotator ablation on rotated ModelNet40 (CPC=50) shows DD3D with PointNet+Rotator at 71.27% vs 62.72% with PCA alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DD3D instead of the strongest listed baseline | Point cloud dataset distillation · Shape classification across ScanObjectNN, ModelNet40, and MVPNet100 (aligned and rotated; CPC=1, 10, 50) | accuracy | Helped |
| Use DD3D rather than GM | Point cloud part segmentation · ShapeNet-part, CPC=1 and CPC=10 | OA: 65.56 -> 73.06; 78.32 -> 80.37 | Helped |
| Use DD3D rather than GM | Point cloud part segmentation · ShapeNet-part, CPC=1 and CPC=10 | Instance IoU: 50.98 -> 60.27; 65.79 -> 66.70 | Helped |
| Use DD3D rather than GM | Point cloud part segmentation · ShapeNet-part, CPC=1 and CPC=10 | Class IoU: 33.96 -> 37.73; 49.88 -> 50.59 | Helped |
| Add PCA canonical-orientation projection (ablation) | PointNet on rotated ModelNet40, CPC=50 · ModelNet40 Rotated | accuracy: 14.75 -> 60.77 | Helped |
| Add the point cloud rotator instead of PCA (ablation) | PointNet on rotated ModelNet40, CPC=50 · ModelNet40 Rotated | accuracy: 60.77 -> 70.13 | Helped |
| Distill with GM rather than using a random coreset | Point cloud segmentation · ShapeNet-part, CPC=1 and CPC=10 | OA: 61.24 -> 65.56; 77.78 -> 78.32 | Helped |
| Train DD3D using 256 or 512 points rather than 1024 points (ablation) | DD3D · ScanObjectNN, ModelNet40, and MVPNet100, CPC=50 | accuracy: ScanObjectNN: 61.27, 60.59, 61.96; ModelNet40: 83.03, 83.59, 83.91; MVPNet100: 39.88, 40.13, 40.61 | No clear effect |
| Generate 1,024-point evaluation clouds after training with lower-resolution inputs (ablation) | DD3D · Point cloud classification | memory usage | Helped |
| Evaluate DD3D-distilled data across architectures rather than only the distillation backbone | DD3D, CPC=50 · ScanObjectNN, ModelNet40, and MVPNet100 | accuracy | Helped |
Surprising findings
- Training on 256-point clouds yields performance similar to 1024-point clouds, and on ScanObjectNN low resolution even slightly outperforms high resolution.
Limitations
- Not applied to scene-level tasks such as object detection due to foreground/background imbalance and continuous box regression.
- Cross-architecture evaluation models do not match PointNet performance, indicating bias from the distillation model.
- Part segmentation limited to CPC=1 and 10 due to GPU memory, and DD3D does not reach full-dataset performance.
Experimental setup
Datasets: ScanObjectNN ModelNet40 MVPNet100 ShapeNet-part
Benchmarks: ScanObjectNN ModelNet40 MVPNet100 ShapeNet-part
Metrics: overall accuracy instance IoU class IoU
Baselines: Random Herding K-Center Gradient Matching Distribution Matching Trajectory Matching
Builds on: gradient matching SIREN PCA vector neurons
Closest papers by meaning
- Dataset Distillation of 3D Point Clouds via Distribution Matching (NeurIPS 2025 · similarity 0.92)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model (AAAI 2025 · similarity 0.86)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.86)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.86)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.86)
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.