Official Repository of CVPR 2019 Paper : RL-GAN-Net: A Reinforcement Learning Agent Controlled GAN Network for Real-Time Point Cloud Shape Completion
https://arxiv.org/abs/1904.12304
This repository is still under constructions.
Requirements:
The packages in my Conda Environment are listed in Requirement_Conda.txt and Requirements_pip.txt files. Only install the ones needed or you can clone the whole environment.
Steps
- Visualize each training and testing step by using visdom.
- Download data from https://github.com/optas/latent_3d_points.
- Process Data with Processdata2.m to get incomplete point cloud
- Train the autoencoder using main.py and save the model
- Generate GFV using pretrained AE using GFV.py and store data
- Train GAN on the generated GFV data by by going into the GAN folder and save model
- Train RL by using pre-trained GAN and AE by running trainRL.py
- Test with Incomplete data by running testRL.py
Credits:
- https://github.com/optas/latent_3d_points
- https://github.com/heykeetae/Self-Attention-GAN
- https://github.com/lijx10/SO-Net (for chamfer distance)
- https://github.com/sfujim/TD3
If you use this work for your projects, please take the time to cite our CVPR paper:
@InProceedings{Sarmad_2019_CVPR,
author = {Sarmad, Muhammad and Lee, Hyunjoo Jenny and Kim, Young Min},
title = {RL-GAN-Net: A Reinforcement Learning Agent Controlled GAN Network for Real-Time Point Cloud Shape Completion},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}