Research map / Point clouds / Point cloud classification
Dataset Distillation of 3D Point Clouds via Distribution Matching
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 a distribution-matching dataset distillation framework for 3D point clouds that sorts per-channel features to fix point-order misalignment (SADM loss) and jointly learns rotation angles for the synthetic objects.
Problem
Dataset distillation for 3D point clouds is largely unexplored; unordered points cause semantic misalignment in feature comparison, and arbitrary object orientations cause rotational variation within classes.
Approach
Compute the MMD-style Gaussian-kernel distribution matching loss on channel-wise descending-sorted point features (SADM, plus a top-feature term), and jointly optimize synthetic point clouds with learnable x/y/z rotation angles, using a randomly initialized PointNet as the feature extractor.
Key results
- At PPC 10 on ModelNet40 the method reaches 80.1% accuracy with PointNet versus 75.6% for DC and 75.2% for DM, and whole-dataset training gives 87.84%.
- Cross-architecture results at PPC 3 show the distilled sets generalize to PointNet++, DGCNN, PointConv and Point Transformer, with the method best on all four datasets in the reported table.
- Under equal training time to DM, the method improves ModelNet40 accuracy at PPC 3 from 62.04% to 70.56%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed SADM-based distillation instead of the strongest listed baseline | PointNet dataset distillation · ModelNet10, PPC 3 | accuracy: 78.0 -> 84.4 | Helped |
| use the proposed SADM-based distillation instead of the strongest listed baseline | PointNet dataset distillation · ModelNet40, PPC 3 | accuracy: 68.5 -> 72.1 | Helped |
| use the proposed SADM-based distillation instead of the strongest listed baseline | PointNet dataset distillation · ShapeNet, PPC 3 | accuracy: 60.1 -> 63.7 | Helped |
| use the proposed SADM-based distillation instead of the strongest listed baseline | PointNet dataset distillation · ScanObjectNN, PPC 3 | accuracy: 27.7 -> 31.6 | Helped |
| add semantic alignment (SA) via sorted feature matching (ablation) | Proposed method · ModelNet10, ModelNet40, ShapeNet, and ScanObjectNN, PPC 3 | accuracy: 75.01 -> 84.42; 59.96 -> 72.08; 53.91 -> 63.74; 20.20 -> 31.84 | Helped |
| add learnable rotation-angle optimization while leaving the synthetic set unoptimized (ablation) | Proposed method · Aligned, mixed, and rotated groups, PPC 3 | accuracy: 65.01 -> 66.22; 59.96 -> 62.17; 20.42 -> 24.89 | Helped |
| add learnable rotation-angle optimization to synthetic-dataset optimization (ablation) | Proposed method · Aligned, mixed, and rotated groups, PPC 3 | accuracy: 73.96 -> 74.35; 71.51 -> 72.08; 29.72 -> 31.84 | Helped |
| replace SADM feature sorting with axis-aligned or Z-order point sorting (ablation) | Proposed method without rotation optimization · ModelNet10, ModelNet40, ShapeNet, and ScanObjectNN, PPC 3 | accuracy: SADM: 83.17, 67.52, 60.71, 27.66; Axis-Aligned: 76.44, 60.85, 53.97, 21.80; Z-order: 76.55, 61.47, 55.17, 21.76 | Mixed |
| replace no rotation augmentation with random rotation augmentation (ablation) | Proposed method · Aligned, mixed, and rotated groups, PPC 3 | accuracy: 65.01 -> 58.92; 59.96 -> 52.56; 20.42 -> 31.42 | Mixed |
| use learnable rotation estimation instead of random rotation augmentation (ablation) | Proposed method · Aligned, mixed, and rotated groups, PPC 3 | accuracy: 58.92 -> 74.35; 52.56 -> 72.08; 31.42 -> 31.84 | Helped |
Surprising findings
- Random rotation augmentation degrades accuracy on aligned datasets (e.g., Aligned 65.01 to 58.92 at PPC 3), while learned rotations help.
- DC with noise initialization fails badly on point clouds (e.g., 5.24% on ShapeNet average), attributed to the large parameter count of PointNet.
Limitations
- SADM does not fully align semantically meaningful regions across different point clouds.
- Rotation estimation adds complexity, especially when datasets are already aligned to canonical axes.
- Evaluated only with PointNet during distillation; the training time comparison is against DM and DC only.
Experimental setup
Datasets: ModelNet10 ModelNet40 ShapeNet ScanObjectNN
Benchmarks: ModelNet10 ModelNet40 ShapeNet ScanObjectNN
Metrics: classification accuracy training time (hours)
Baselines: Random selection Herding K-center DM DC PCC MTT DataDAM Z-order sorting
Builds on: Dataset Condensation with Distribution Matching (DM) Dataset Condensation with Gradient Matching (DC) PointNet Maximum Mean Discrepancy
Closest papers by meaning
- Point Cloud Dataset Distillation (ICML 2025 · similarity 0.92)
- Parameterization-Based Dataset Distillation of 3D Point Clouds through Learnable Shape Morphing (ICLR 2026 · similarity 0.91)
- Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation (NeurIPS 2025 · similarity 0.88)
- DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features (CVPR 2025 · similarity 0.87)
- Hyperbolic Dataset Distillation (NeurIPS 2025 · similarity 0.87)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (NeurIPS 2025 · 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.