6.7 KiB
Gaussian Haircut: Human Hair Reconstruction with Strand-Aligned 3D Gaussians
This repository contains the official implementation of Gaussian Haircut, a strand-based human hair reconstruction method from monocular video.
Overview
The reconstruction process includes the following main stages:
-
Preprocessing Stage
- Video frame extraction and organization
- COLMAP camera reconstruction
- Hair and body segmentation
- Image quality assessment and filtering
- Orientation map calculation
- Facial keypoint detection
- FLAME head model fitting
-
Reconstruction Stage
- 3D Gaussian reconstruction
- FLAME mesh fitting
- Scene cropping and optimization
- Hair strand reconstruction
-
Visualization Stage
- Export reconstructed strands
- Blender rendering visualization
- Generate result video
Expected output:
[your_scene_folder]/
├── raw.mp4 # Input video
├── 3d_gaussian_splatting/ # 3D Gaussian reconstruction results
├── flame_fitting/ # FLAME head model fitting results
├── strands_reconstruction/ # Hair strand reconstruction intermediate results
├── curves_reconstruction/ # Final hair strand results
└── visualization/ # Rendering results and video
入门指南
Linux 平台
-
安装 CUDA 11.8
按照 https://developer.nvidia.com/cuda-11-8-0-download-archive 上的说明进行操作。
确保:
- PATH 包含 <CUDA_DIR>/bin
- LD_LIBRARY_PATH 包含 <CUDA_DIR>/lib64
该环境仅在此 CUDA 版本下进行了测试。
-
安装 Blender 3.6 以创建股线可视化
按照 https://www.blender.org/download/lts/3-6 上的说明进行操作。
-
克隆仓库并运行安装脚本
git clone git@github.com:eth-ait/GaussianHaircut.git cd GaussianHaircut chmod +x ./install.sh ./install.sh
Windows 平台
-
安装 CUDA 11.8
- 从 https://developer.nvidia.com/cuda-11-8-0-download-archive 下载并安装
- 默认安装路径:C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8
- 确保CUDA版本与系统兼容
-
安装 Blender 3.6
- 从 https://www.blender.org/download/lts/3-6 下载并安装
- 默认安装路径:C:\Program Files\Blender Foundation\Blender 3.6
-
安装 Visual Studio 2019 Build Tools
- 从 https://visualstudio.microsoft.com/vs/older-downloads/ 下载并安装
- 选择"C++构建工具"工作负载
- 默认安装路径:C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools
-
安装 COLMAP
- 从 https://github.com/colmap/colmap/releases 下载并安装
- 下载CUDA版本的COLMAP (例如:COLMAP-3.8-windows-cuda.zip)
- 解压到不含空格的路径 (例如:C:\COLMAP)
- 将COLMAP目录添加到系统PATH:
- 打开"系统属性" > "环境变量"
- 在"系统变量"中找到"Path"
- 点击"编辑" > "新建"
- 添加COLMAP目录路径
- 点击"确定"保存
- 重启终端使PATH生效
-
安装 7-Zip
- 从 https://7-zip.org/ 下载并安装
- 将7-Zip安装目录添加到系统PATH:
- 打开"系统属性" > "环境变量"
- 在"系统变量"中找到"Path"
- 点击"编辑" > "新建"
- 添加7-Zip安装目录(默认为C:\Program Files\7-Zip)
- 点击"确定"保存
- 重启终端使PATH生效
-
下载预训练模型和资源
git clone https://github.com/Jeffreytsai1004/GaussianHairCut cd GaussianHairCut # 在PowerShell中运行: # 脚本会自动安装gdown并下载所需资源 .\download_resource.bat
注意:
- 下载过程可能需要几分钟到几十分钟,取决于网络速度
- 如果下载失败,可以重新运行脚本
- 确保有稳定的网络连接
-
克隆仓库并运行安装脚本
git clone https://github.com/Jeffreytsai1004/GaussianHairCut cd GaussianHairCut # 先下载所需要的资源 .\download_resource.bat # 运行安装脚本 .\install.bat # 运行重建脚本 .\run.bat
重建
-
录制单目视频
- 参考项目页面上的示例视频
- 录制要求:
- 拍摄对象应缓慢转动头部,确保捕捉到所有角度
- 保持头发和面部清晰可见
- 避免快速移动导致的运动模糊
- 保持光照条件稳定
- 建议视频长度:10-20秒
- 建议分辨率:1920x1080或更高
-
设置重建场景的目录
创建场景目录结构:
[your_scene_folder]/ └── raw.mp4 # 您录制的视频文件
注意:
- DATA_PATH 应指向包含 raw.mp4 的目录
- 目录路径不应包含空格或特殊字符
- 脚本会在此目录下自动创建所需的子目录
- 确保有足够的磁盘空间(建议至少20GB)
-
运行脚本
Linux:
export PROJECT_DIR="[/path/to/]GaussianHaircut"
export BLENDER_DIR="[/path/to/blender/folder/]blender"
DATA_PATH="[path/to/scene/folder]" ./run.sh
Windows:
# 在CMD中运行:
set PROJECT_DIR=[path\to\]GaussianHaircut
set DATA_PATH=[path\to\scene\folder]
run.bat
# 或在PowerShell中运行:
set PROJECT_DIR=[path\to\]GaussianHaircut
set DATA_PATH=[path\to\scene\folder]
.\run.bat
该脚本执行数据预处理、重建以及使用 Blender 进行最终可视化。使用 Tensorboard 查看中间可视化结果。
许可证
此代码基于 3D Gaussian Splatting 项目。有关条款和条件,请参阅 LICENSE_3DGS。其余代码根据 CC BY-NC-SA 4.0 分发。
如果此代码对您的项目有帮助,请引用以下论文。
引用
@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}
}
链接
- 3D Gaussian Splatting
- Neural Haircut: FLAME 拟合管线、股线先验和发型扩散先验
- HAAR: 头发上采样
- Matte-Anything: 头发和身体分割
- PIXIE: FLAME 拟合的初始化
- Face-Alignment, OpenPose: 用于 FLAME 拟合的关键点检测