TreeGAN

This repository TreeGAN is for 3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions paper accepted on ICCV 2019

---

[ Paper ]

3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions
(Dong Wook Shu*, Sung Woo Park*, Junseok Kwon)

---

[Network]

TreeGAN network consists of "TreeGCN Generator" and "Discriminator".

For more details, refer our paper.

---

[Results]

• Multi Class Generation.
  

• Single Class Generation.
  

• Single Class Interpolation.
  

---

[Frechet Pointcloud Distance]

• This FPD version is used pretrained PointNet.

• This FPD version is for ShapeNet-Benchmark dataset from A Scalable Active Framework for Region Annotation in 3D Shape Collections.

• Our pretrained PointNet-FPD version use only subset of official ShapeNet dataset to get PointNet classification performance higher than 95%.

• We recommend to compose pointclouds sampled uniformly from those of ShapeNet-Benchmark dataset for training.

• We evaluate FPD scores using 5000 samples obtained from fixed trained model with best performances.

• FPD evaluations have to use pre_statistics file for each class or all class version.

• We just provide intermediate pretrained checkpoints and generated samples having fine scores when they are trained in about 1000 epochs.

---

[Citing]

@InProceedings{Shu_2019_ICCV,
               author = {Shu, Dong Wook and Park, Sung Woo and Kwon, Junseok},
               title = {3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions},
               booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
               month = {October},
               year = {2019}}

[Setting]

This project was tested on Windows 10 / Ubuntu 16.04 Using conda install command is recommended to setting.

Packages

• Python 3.6
• Numpy
• Pytorch 1.0
• visdom
• Scipy 1.2.1
• Pillow

---

[Arguments]

In our project, arguments.py file has almost every parameters to specify for training.

For example, if you want to train, it needs to specify dataset_path argument.