The source code of the networks for our paper "Dynamic hair modeling from monocular videos using deep neural networks" (SIGGRAPH ASIA 2019)
We propose a novel framework for dynamic hair modeling from monocular videos. We use two networks HairSpatNet and HairTempNet to separately predict hair geometry and hair motion. The entire framework is as follows:
- For HairSpatNet, we removed instance normalization and the discriminator to speed up training process and reduce memory cost. We found that the fine-grained details imposed by the discriminator would be obliterated by the space-time optimization afterwards.
- For motion prediction, we redesigned a network named HairWarpNet to directly predict flow based on the 3D fields (similar to the regression of optical flow). It is more reasonable and achieves better results than HairTempNet.
- There are more designs of toVoxel modules.
- You can check other research directions in folder OtherResearch.
- Linux
- Python 3.6
- NVIDIA GPU + CUDA 10.0 + cuDNN 7.5
- tensorflow-gpu 1.13.1
- Conda installation:
# 1. Create a conda virtual environment. conda create -n dhair python=3.6 -y conda activate dhair # 2. Install dependency pip install -r requirement.txt
- You can run the scripts in the Script folder to train/test your models.
If you find this useful for your research, please cite the following paper.
@article{yang2019dynamic,
title={Dynamic hair modeling from monocular videos using deep neural networks},
author={Yang, Lingchen and Shi, Zefeng and Zheng, Youyi and Zhou, Kun},
journal={ACM Transactions on Graphics (TOG)},
volume={38},
number={6},
pages={1--12},
year={2019},
}