Point Cloud GAN

This repository implements a modified version of Point Cloud GAN (ICLR'19 Workshop) which achieves performance comparable to the SetVAE in point cloud generation.

Installation

• Set up and activate conda environment.

conda env create -f environment.yml
conda activate pcgan

• Compile CUDA extensions.

sh scripts/install.sh

• Download ShapeNet dataset and trained checkpoints.

sh scripts/download.sh

Training

You can train using train.py or provided scripts.

# Train using CLI
python train.py --name NAME --cate airplane
# Train using provided settings
sh scripts/train_shapenet_airplane.sh

Testing

You can evaluate checkpointed models using test.py or provided scripts.

# Test user specified checkpoint using CLI
python test.py --ckpt_path CKPT_PATH --cate airplane
# Test provided checkpoints
sh scripts/test_shapenet_airplane.sh

Metrics

Table below shows final metrics for SetVAE and our model (MMD-CD is scaled by 10³ and MMD-EMD, COV, 1-NNA by 10²). SetVAE is trained for 8000 epochs and our model is trained for 2000 epochs.

Category	Model	MMD(↓) CD	MMD(↓) EMD	COV(↑) CD	COV(↑) EMD	1-NNA(↓) CD	1-NNA(↓) EMD
Airplane	SetVAE	0.199	3.07	43.45	44.93	75.31	77.65
	Ours	0.224	3.45	38.27	36.79	-	-