This code was maded based on PointNet.(repository : https://github.com/charlesq34/pointnet) And I used chamfer distance + repulsion distance as a point cloud loss function. Repulsion distance was used for uniform distribution of the point cloud. (paper : https://arxiv.org/abs/1801.06761) Also I used the autoencoder model as a generator of GAN.

To use tf_grouping type make. Also to use tf_nn_distance change makefile_dist to makefile and type make. Before train, you must download shapenet datasets and make data dir and copy or move the shapenet datasets data dir. (Download datasets)

To train model type "python train.py".

If you want to see the scattered point cloud, you must download visdom and type "python -m visdom.server".

Here, we will be able to use the GAN model to reconstruct a damaged 3D point cloud by classifying it using feature distance to find out what shape of a point cloud it is.

This is a screenshot of the test data applied to the model.