SynthHuman Dataset
收藏资源简介:
SynthHuman数据集包含大约300,000张合成人类的图像,带有前景alpha掩码、绝对深度、表面法线和相机内参的真实标注。数据集分为三个相机场景:面部、上半身和全身,每个场景大约有100,000张图像。数据是通过最新的合成数据生成流程生成的,该流程已用于创建多个数据集:Face Synthetics、SimpleEgo和SynthMoCap。由于基于图形的渲染流程,真实标注是像素级的,具有完美准确性。
The SynthHuman dataset contains approximately 300,000 synthetic human images, with ground-truth annotations including foreground alpha masks, absolute depth maps, surface normals, and camera intrinsic parameters. The dataset is divided into three camera scenarios: face, upper body, and full body, with roughly 100,000 images per scenario. The data is generated via a state-of-the-art synthetic data generation pipeline that has been used to create multiple datasets including Face Synthetics, SimpleEgo, and SynthMoCap. Owing to the graphics-based rendering pipeline, the ground-truth annotations are pixel-level and feature perfect accuracy.
DAViD数据集概述
数据集基本信息
- 名称:SynthHuman Dataset
- 数据量:约300,000张合成人类图像
- 标注类型:前景alpha掩膜、绝对深度、表面法线和相机内参
- 数据场景:
- 面部:约100,000张
- 上半身:约100,000张
- 全身:约100,000张
- 数据格式:
rgb_0000000.png:RGB图像alpha_0000000.png:前景alpha掩膜depth_0000000.exr:绝对深度图像(厘米单位)normal_0000000.exr:表面法线图像(XYZ)cam_0000000.txt:相机内参矩阵
数据下载与使用
- 下载方式:通过
download_data.py脚本下载 - 数据分块:60个压缩文件,每个约8.75GB,总下载量约330GB
- 可视化工具:
visualize_data.py脚本 - 许可证:CDLA-2.0许可证
发布模型
- 任务与模型:
- 许可证:MIT许可证
演示运行
- 支持任务:相对深度估计、软前景分割、表面法线估计
- 运行方式:
- 使用多任务模型
- 使用单独模型
- 依赖安装:
pip install -r requirement.txt
引用
如需使用SynthHuman数据集或DAViD模型,请引用: bibtex @misc{saleh2025david, title={{DAViD}: Data-efficient and Accurate Vision Models from Synthetic Data}, author={Fatemeh Saleh and Sadegh Aliakbarian and Charlie Hewitt and Lohit Petikam and Xiao-Xian and Antonio Criminisi and Thomas J. Cashman and Tadas Baltrušaitis}, year={2025}, eprint={2507.15365}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2507.15365}, }




