CPCGAN: A Controllable 3D Point Cloud Generative Adversarial Network with Semantic Label Generating

[Paper] [Code] [Poster] [Video]
The official code repository for our AAAI 2021 paper CPCGAN: A Controllable 3D Point Cloud Generative Adversarial Network with Semantic Label Generating.

System requirements

Our model has been tested with PyTorch 1.7.0 and CUDA 11.0

Usage

Step1 Setup environment

# Install DLNest
pip install git+https://github.com/SymenYang/DLNest.git
# Install other libs
pip install tensorboard
pip install sklearn
# Set path in config
bash BeforeTrain.sh

Step2 Place the dataset. Please locate the shapenet-partseg dataset as follow:

|--CPCGAN
   |--AnalyzeScripts
   |--Datas
      |--train_data
         |--02691156
            |--000001.pts
            |--...
         |--02773838
         |--...
      |--train_label
         |--02691156
            |--000001.seg
            |--...
         |--02773838
         |--...
   |--Dataset
   |--Model
   |--common_config.json
   |--dataset_config.json
   |--freq_config.json
   |--model_config.json
   |--root_config.json
   |--LifeCycle.py

Step3 Preprocess the datas. This may takes 1 or 2 hours due to the K-Means calculations.

cd DataPreprocessing
bash DataPreprocessing.sh

Step4 Train CPCGAN using DLNest.

python -m DLNest.Run -c <Absolute path to the root_config.json> -f <Absolute path to the freq_config.json>

or run this command in DLNest's shell client:

run -c <Absolute path to the root_config.json> -f <Absolute path to the freq_config.json>

The configs,backup codes,checkpoints and training outputs will be restored in the Saves/<A timestamp> directory.
More details about DLNest and DLNest-based projects please refer to DLNest (Only Chinease docs yet.)
Step5 Tests

1. Test in FPD metric.

python -m DLNest.Analyze -r <Absolute path to the save dir(Saves/<A timestamp>)> -s <Absolute path to AnalyzeScripts/get_FPD.py> -c <The best epoch>>

or run these commands in DLNest's shell client:

analyze -r <Absolute path to the save dir(Saves/<A timestamp>)>
watch <The task ID of this analyze process>
runExp get_FPD

2. Generate a sample.

python -m DLNest.Analyze -r <Absolute path to the save dir(Saves/<A timestamp>)> -s <Absolute path to AnalyzeScripts/gen_a_sample.py> -c <The best epoch>>

or run this command in DLNest's shell client(After analyze process is started in above test):

runExp gen_a_sample 

3. Control the generation.
   CPCGAN can control the generated shape by modifying the structure point cloud. Please add the modifying codes to the modify_pc_0 function in AnalyzeScripts/gen_from_spc_and_z.py. Then run this command:

python -m DLNest.Analyze -r <Absolute path to the save dir(Saves/<A timestamp>)> -s <Absolute path to AnalyzeScripts/gen_from_spc_and_z.py> -c <The best epoch>>

or run this command in DLNest's shell client(After analyze process is started in above test):

runExp gen_from_spc_and_z

Citation

@article{Yang_Wu_Zhang_Jin_2021, 
   title={CPCGAN: A Controllable 3D Point Cloud Generative Adversarial Network with Semantic Label Generating}, 
   volume={35}, 
   url={https://ojs.aaai.org/index.php/AAAI/article/view/16425}, 
   number={4}, 
   journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
   author={Yang, Ximing and Wu, Yuan and Zhang, Kaiyi and Jin, Cheng}, 
   year={2021}, 
   month={May}, 
   pages={3154-3162}
}