Gaussian Haircut: Human Hair Reconstruction with Strand-Aligned 3D Gaussians

Paper | Project Page

This repository contains an official implementation of Gaussian Haircut, a strand-based hair reconstruction approach for monocular videos.

Getting started

1. Install CUDA 11.8

   Follow the instructions on https://developer.nvidia.com/cuda-11-8-0-download-archive.

   Make sure that

   • PATH includes <CUDA_DIR>/bin
   • LD_LIBRARY_PATH includes <CUDA_DIR>/lib64

   The environment was tested only with this CUDA version.

2. Install Blender 3.6 in order to create strand visualizations

   Follow instructions on https://www.blender.org/download/lts/3-6.

3. Close the repo and run the install script

   git clone git@github.com:eth-ait/GaussianHaircut.git
   cd GaussianHaircut
   chmod +x ./install.sh
   ./install.sh

Reconstruction

1. Record a monocular video

   Use examples on the project page as references and introduce as little motion blur as possible.

2. Setup a directory for the reconstructed scene

   Put the video file in it and rename it to raw.mp4

3. Run the script

   export PROJECT_DIR="[/path/to/]GaussianHaircut"
   export BLENDER_DIR="[/path/to/blender/folder/]blender"
   DATA_PATH="[path/to/scene/folder]" ./run.sh

   The script performs data pre-processing, reconstruction, and visualization.

License

This code is based on the 3D Gaussian Splatting project. For terms and conditions, please refer to LICENSE_3DGS. The rest of the code is distributed under CC BY-NC-SA 4.0.

If this code is helpful in your project, cite the papers below.

Citation

@inproceedings{zakharov2024gh,
   title = {Human Hair Reconstruction with Strand-Aligned 3D Gaussians},
   author = {Zakharov, Egor and Sklyarova, Vanessa and Black, Michael J and Nam, Giljoo and Thies, Justus and Hilliges, Otmar},
   booktitle = {European Conference of Computer Vision (ECCV)},
   year = {2024}
} 

Links

• 3D Gaussian Splatting

• Neural Haircut: FLAME fitting pipeline, strand prior and hairstyle diffusion prior

• HAAR: hair upsampling

• Matte-Anything: hair and body segmentation

• PIXIE: initialization for FLAME fitting

• Face-Alignment, OpenPose: keypoints detection for FLAME fitting