[NeurIPS 2025] Dataset Distillation of 3D Point Clouds via Distribution Matching

Figure: Overall framework of the proposed dataset distillation method.

---

Introduction

Large-scale 3D point cloud datasets are essential for training deep neural networks but impose heavy computational burdens. We propose a distribution matching-based dataset distillation framework that jointly optimizes both geometric structures and orientations of synthetic 3D point clouds.

Key contributions:

• Semantically Aligned Distribution Matching (SADM): Introduces feature sorting within each channel to mitigate semantic misalignment caused by unordered point indices.
• Orientation-Aware Optimization: Learns optimal rotations of synthetic 3D objects, reducing intra-class variability and improving feature alignment.
• Superior Generalization: Achieves strong cross-architecture generalization across PointNet, PointNet++, DGCNN, and Point Transformer.

Our method consistently outperforms existing dataset distillation and coreset selection approaches on benchmarks such as ModelNet10, ModelNet40, ShapeNet, and ScanObjectNN.

---

We visualize the distilled synthetic datasets compared with baseline methods (DC, DM) on ModelNet40. Our method preserves the semantic structure of 3D objects (e.g., edges, corners, class-specific shapes), while existing methods tend to collapse to noisy or less informative patterns.

Figure: Comparison of synthetic datasets distilled by DC(row1), DM(row2), and our(row3) method.

News

• 2025-09-24: Our paper has been accepted to NeurIPS 2025 🎉
• 2025-05-29: Preprint released on arXiv.

Quickstart

⚠️ The code will be released upon paper publication. Please stay tuned!

git clone https://github.com/donguk071/SADM.git

Citation

If you find our work useful, please cite:

@article{yim2025dataset,
    title={Dataset Distillation of 3D Point Clouds via Distribution Matching},
    author={Yim, Jae-Young and Kim, Dongwook and Sim, Jae-Young},
    journal={arXiv preprint arXiv:2503.22154},
    year={2025},
    url={https://arxiv.org/abs/2503.22154}
}