Neural Octahedral Field

The implementation of preprint Neural Octahedral Field: Octahedral prior for simultaneous smoothing and sharp edge regularization

WARNING This is a research repo with limited code quality. We have only tested it on Linux (Ubuntu 22.04 and EndeavorOS).

Environment Setup

We first create a new environment

conda create -n octa python=3.10 -y
conda activate octa

Follow instructions to install JAX and PyTorch (cpu), e.g.

pip install -U "jax[cuda12]"
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Then install libigl and the rest dependencies by

python -m pip install libigl
pip install -r requirements.txt

(Optional) We also have some CPP bindings referred as frame_field_utils in code. It is subject to difference Licenses and is not required for the main results

Reconstruction

Run

python run_recon.py --config configs/octa_hessian.json --model /path/to/target_pointcloud.ply

or

python run_recon.py --config configs/octa_hessian_noisy.json --model /path/to/target_noisy_pointcloud.ply

based on noise level

See metric folder for datasets and compared methods