3DDP

Official implementation of "Parameterization-Based Dataset Distillation of 3D Point Clouds through Learnable Shape Morphing" (ICLR 2026).

[Paper]

We synthesize a small set of point clouds that faithfully preserves the information of a full training set, enabling networks trained only on the distilled data to match or approach full-data accuracy.

---

Method overview

Adaptive 3D Shape Morphing represents each class with anchor chunks that are linearly combined via learned softmax weights to produce morphed training samples, drastically reducing the number of stored parameters while expanding effective training diversity.

The distillation objective matches per-partition intermediate features between real and synthetic data, weighted by a uniformity score that penalises unevenly distributed chunks (fReshape_nodup loss).

---

Installation

Requires a machine with a CUDA 12.1-compatible GPU and conda.

git clone git@github.com:yimjae0/point-distill.git
cd point-distill
bash install.sh

install.sh creates a conda environment called point-distill with Python 3.10, PyTorch 2.2.2 + CUDA 12.1, and builds the two required CUDA extensions (pointops, emd_).

---

Data preparation

Download the datasets and set environment variables to point to them before running:

Dataset	Environment variable	Default path
ModelNet40 / ModelNet10	MODELNET40_ROOT	/root/dataset/ModelNet40
ScanObjectNN	SONN_ROOT	/root/dataset/ScanObjectNN/main_split_nobg
ShapeNet	SHAPENET_ROOT	/root/dataset/ShapeNetv2/PointCloud
OmniObject3D	OMNI_ROOT	/root/dataset/OmniObject3D

Example:

export MODELNET40_ROOT=/data/ModelNet40

---

Reproducing Table 1

conda activate point-distill
bash scripts/run_table1.sh 0   # 0 = GPU id

This runs all 15 experiments (5 datasets × IPC ∈ {1, 3, 10}) sequentially on a single GPU and saves results to ./result/.

Key hyperparameters

Argument	Description	Default
--ipc	Synthetic samples per class	1
--npoints	Points per anchor chunk	255 (252 for ModelNet10)
--num_morph	Morphed samples per IPC slot	4 (16 for ModelNet10)
--Iteration	Distillation iterations	2000
--init	Initialisation strategy (real / noise)	real
--num_eval	Evaluation runs to average	10
--epoch_eval_train	Epochs for evaluation training	500

Single experiment

conda activate point-distill
export PYTHONPATH="$(pwd)/extensions:$PYTHONPATH"

python scripts/train.py \
  --dataset MODELNET40 \
  --model PointNet \
  --ipc 1 \
  --npoints 255 \
  --num_morph 4 \
  --Iteration 2000 \
  --num_eval 10 \
  --epoch_eval_train 500 \
  --init real \
  --batch_real 8 \
  --batch_train 8 \
  --lr_net 0.01

---

Repository structure

point-distill/
├── install.sh                   # one-shot environment setup
├── scripts/
│   ├── train.py                 # main distillation script
│   └── run_table1.sh            # Table 1 reproduction
├── src/point_distill/
│   ├── data/datasets.py         # dataset loaders
│   ├── distill/losses.py        # M3D loss
│   ├── models/factory.py        # backbone factory
│   └── ops/
│       ├── pc_ops.py            # FPS, EMD alignment, uniformity score
│       └── training.py          # evaluation loop
└── extensions/
    ├── pointops/                # CUDA: KNN, ball query, FPS, grouping
    └── emd_/                    # CUDA: Earth Mover's Distance

---

Supported backbones

PointNet, PointNetPlusPlus, DGCNN, PointConvDensityClsSsg, PointTransformerCls

Pass via --model.

---

Citation

@inproceedings{pointdistill2026,
  title     = {Parameterization-Based Dataset Distillation of 3D Point Clouds through Learnable Shape Morphing},
  booktitle = {ICLR},
  year      = {2026},
}