/Temporal-Pointcloud-Upsampling-GAN

Primary LanguagePythonMIT LicenseMIT

Temporal-Pointcloud-Upsampling-GAN

Code repo for TPU-GAN (paper)

Requirement

Other packages can be installed through: conda env create -f environment.yml

Optional (for post analysis or I/O, not required for training):

The code is tested under Linux Ubuntu 18.04 with CUDA 10.2.

Dataset

We use DFSPH solver from SPLISHSPLASH (https://github.com/InteractiveComputerGraphics/SPlisHSPlasH) to generate simulation data. The geometry assets and Python script for simulation are taken and tweaked from CConv, Benjamin Ummenhofer et al. under CDLA-Permissive-1.0 and MIT license respectively.

To generate fluid data, first install SPLISHSPLASH following: https://splishsplash.readthedocs.io/en/latest/build_from_source.html (with Python bindings turned on), and set the path to binary at line 3 of splishsplash_config.py under fluid_data_generation folder. Download geometry assets from link and put them under the folder fluid_data_generation/models.

Then run SPH simulation by:

cd fluid_data_generation
python sim_fluid_sequence.py

And process the simulation data (in bgeo format) into npz data. This step requires Partio installed.

python process_training_data.py

We also provide generated data at link.

The MSR-Action dataset is taken from MeteorNet: https://github.com/xingyul/meteornet/tree/master/action_cls.

Training and inference

The training scripts and pretrained models are under subfolder train_fluid and train_action. Please refer to them for more details. Before training, create a folder named data (under base directory, i.e. Temporal-Pointcloud-Upsampling-GAN/data) and put collected data under this folder with following structure:

data
├── train_data_0.025_fine (for fluids)
├── test_data_0.025_fine (for fluids)
├── bunny (for fluid demo)
├── MSR-Action3D (for action)

Visualization and rendering

We use Maya with Partio plugin to load and render the results (in .bgeo format).

Acknowledgement

The code in this project is partially based on or inspired from following projects.