HairStep: Transfer Synthetic to Real Using Strand and Depth Maps for Single-View 3D Hair Modeling (CVPR2023 Highlight) [Projectpage]

Dataset

All data of HiSa & HiDa is hosted on Google Drive:

Path	Files	Format	Description
HiSa_HiDa	12,503		Main folder
├ img	1,250	PNG	Hair-centric real images at 512×512
├ seg	1,250	PNG	Hair masks
├ body_img	1,250	PNG	Whole body masks
├ stroke	1,250	SVG	Hair strokes (vector curves) manually labeled by artists
├ strand_map	1,250	PNG	Strand maps with body mask
├ camera_param	1,250	NPY	Estimated camera parameters for orthogonal projection
├ relative_depth	2,500		Folder for the annotation of the ordinal relative depth
├ pairs	1,250	NPY	Pixel pairs randomly selected from adjacent super-pixels
├ labels	1,250	NPY	Ordinal depth labels for which pixels are closer in pixel pairs
├ dense_depth_pred	2,500		Folders for dense depth maps generated by our domain adaptive depth estimation method
├ depth_map	1,250	NPY	Nomalized depth maps (range from 0 to 1, the closer the bigger)
├ depth_vis	1,250	PNG	Visulization of depth maps
├ split_train.json	1	JSON	Split file for training
└ split_test.json	1	JSON	Split file for testing

The HiSa & HiDa dataset and pre-trained checkpoints based on it are available for non-commercial research purposes only. All real images are collected from the Internet. Please contact Yujian Zheng and Xiaoguang Han for questions about the dataset.

Installation

git clone --recursive https://github.com/GAP-LAB-CUHK-SZ/HairStep.git

cd HairStep

conda env create -f environment.yml
conda activate hairstep

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html

pip install -r requirements.txt

cd external/3DDFA_V2
sh ./build.sh
cd ../../

Code is tested on torch1.9.0, CUDA11.1, Ubuntu 20.04 LTS.

Single-view 3D Hair Reconstruction

Put collected and cropped potrait images into ./results/real_imgs/img/.

Download the checkpoint of SAM and put it to ./checkpoints/SAM-models/.

Download checkpoints of 3D networks and put them to ./checkpoints/recon3D/.

CUDA_VISIBLE_DEVICES=0 python -m scripts.img2hairstep
CUDA_VISIBLE_DEVICES=0 python scripts/get_lmk.py
CUDA_VISIBLE_DEVICES=0 python -m scripts.opt_cam
CUDA_VISIBLE_DEVICES=0 python -m scripts.recon3D

Results will be saved in ./results/real_imgs/.

TODO

Share the HiSa & HiDa datasets
Release the code for converting images to HairStep
Release the code for reconstructing 3D strands from HairStep
Release the code for computing metrics HairSale & HairRida
Release the code for training and data pre-processing

(within a few days before the end of Feb., sorry for the delay)

Note: A more compact and efficient sub-module for 3D hair reconstruction has been released, which has comparable performance to NeuralHDHair* reported in the paper.

The original hair matting approach is provided by Kuaishou Technology, which cannot be released. The substitute method based on SAM fails sometimes.

Citation

Please cite our paper as below if you find this repository is helpful:

@inproceedings{zheng2023hairstep,
  title={Hairstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling},
  author={Zheng, Yujian and Jin, Zirong and Li, Moran and Huang, Haibin and Ma, Chongyang and Cui, Shuguang and Han, Xiaoguang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12726--12735},
  year={2023}
}

Acknowledgements

This repository is based on some excellent works, such as HairNet, PIFu, 3DDFA_V2, SAM and Depth-in-the-wild. Many thanks.