• PointOT
• • Installation
  • Dataset
  • Usage
  • • Training
    • Generation and Evaluation
• Futher Information

PointOT

This repository contains the code to reproduce the results from the paper
PointOT: Interpretable Geometry-Inspired Point Cloud Generative Model via Optimal Transport.

You can find detailed usage instructions for training your own models and using pretrained models below.

If you find our code or paper useful, please consider citing

@inproceedings{PointOT,
    title = {{PointOT}: Interpretable Geometry-Inspired Point Cloud Generative Model via Optimal Transport},
    author = {Ruonan Zhang and Jingyi Chen and Wei Gao and Ge Li and Thomas H. Li},
    booktitle = {IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY},
    year = {2022}
}

Installation

First you have to make sure that you have all dependencies in place.
The simplest way to do so, is to use anaconda.

You can create an anaconda environment called pointot using

conda env create -f pointot.yaml
conda activate pointot

for emd-loss:

cd ./models/emd/
python setup.py install

Dataset

This dataset is already put in cloud brain, named: shape_net_core_uniform_samples_2048.zip
or you can download here: https://git.openi.org.cn/attachments/62228286-8ddf-42f7-bfe8-6a420cbe1576?type=0

Other methods to download:
Download data ShapeNetCore with (area) uniform sampling from https://github.com/optas/latent_3d_points.

the dataset is located in : /data/rayna/datasets/shape_net_core_uniform_samples_2048/

for better distinct classes, in our dataset, they are rearranged as follows:

-shape_net_core_uniform_samples_2048
--four
----02691156
----02958343
----03001627
----04379243
--eight
----02691156
----02828884
----02933112
----02958343
----03001627
----03211117
----04256520
----04379243
--airplane
----02691156
--car
----02958343
--chair
----03001627
--table
----04379243

Usage

When you have installed all binary dependencies and obtained the preprocessed data, you are ready to run our pretrained models and train new models from scratch.

Training

Step 1: Train point cloud auto-encoder, fist open visdom,

python -m visdom.server

then, run

bash trainAE.sh

Note: the trained AE models are stored in ./ckpts/AE_ckpts/ folder.

The point cloud AE is establised based on RL-GAN-Net:

Step 2: Train probability distribution transformation, run

bash trainOT.sh

Note: the trained OT models are stored in ./ckpts/OT/ folder.

Generation and Evaluation

Some metrics (KL JSD F-score Sink) and generation results part, please run,

bash test_trainOT.sh

action: generate, for generating the new latent distribution
action: decode_feature, for testing the results of trained datasets
action: decode_test, for generating the new point clouds and test results

other evaluations on measurements JSD\MMD\1-NNA, you can run:

bash eval_metrics.sh

Note: the environments problems please refer to Diffusion Probabilistic Models for 3D Point Cloud Generation for more help. We run this in the environment of eval.yaml

Futher Information

Please also check out the following codes:

• RL-GAN-Net
• pyOMT: A Pytorch implementation of Adaptive Monte Carlo Optimal Transport Algorithm