Awesome

Image Based Reconstruction of Liquids from 2D Surface Detections

Still under construction!

dataset: https://drive.google.com/drive/folders/1b2TIdIdH4HRRcTMPFetI_I1mHnlXQ-LY?usp=sharing

arxiv: https://arxiv.org/abs/2111.11491

Environment set up:

Tested with Python 3.8, PyTorch 1.7.1, CUDA 11.0 and PyTorch3D 0.4.0

conda create --name liquid_rec python=3.8
conda activate liquid_rec
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d=0.4.0=py38_cu110_pyt171 -c pytorch3d

Requires SPNet (original github is: https://github.com/cschenck/SmoothParticleNets) for solving collision constraint. The following fork updated the original SPNet for PyTorch 1.7.1

mkdir liquid_reconstruction
cd liquid_reconstruction
gh repo clone bango123/SmoothParticleNets
cd SmoothParticleNets
python setup.py install

Install dependencies for this repo (Open3d for visualization and trimesh & mesh-to-sdf to generate SDF's from mesh):

pip install open3d
pip install trimesh
pip install mesh-to-sdf

Install this repo

cd ..
conda install -c anaconda sympy
gh repo clone ucsdarclab/liquid_reconstruction

Run Example Code

While not perfectly tuned, an example of PBF simulation (https://mmacklin.com/pbf_sig_preprint.pdf) is presnted in simulateBox.py. This is to show how the functional components in differentiableFluid.py behave.

python simulateBox.py