---
title: HumanNet
canonical_url: "https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet"
md_url: "https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet.md"
repository: DAGroup-PKU/HumanNet
last_updated: 2026-11-11
license: other
storage_size: "90 TB"
downloads: 6864493
stars: 23
---

# HumanNet

> HumanNet - DAGroup-PKU 在 ModelScope 开源的数据集。面向具身智能的百万小时人类中心视频语料 北京大学 DAGroup · SimpleSilicon Innovation Team 100 万小时整体语料 · 20 万小时 Ego 完整发布 · 1 万小时 HF 子集

DAGroup-PKU/HumanNet 是 ModelScope 魔搭社区上的数据集，存储大小 90 TB，采用 other 许可。

- **Repository**: DAGroup-PKU/HumanNet
- **License**: other
- **Storage size**: 90 TB
- **Downloads**: 6864493
- **Stars**: 23
- **Last updated**: 2026-11-11

Source: https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet

---

<p align="center"><img src="https://raw.githubusercontent.com/DAGroup-PKU/HumanNet/main/assets/branding/humannet-header-shadow.svg" width="1000" alt="HumanNet · 北京大学 DAGroup" /></p>

<h3 align="center">面向具身智能的百万小时人类中心视频语料</h3>
<p align="center">北京大学 DAGroup · SimpleSilicon Innovation Team</p>
<p align="center"><strong><span style="color:#990100;font-size:24px">100 万小时整体语料 · 20 万小时 Ego 完整发布 · 1 万小时 HF 子集</span></strong></p>

[HumanNet 论文](https://arxiv.org/abs/2605.06747) · [项目主页](https://dagroup-pku.github.io/HumanNet/) · [GitHub](https://github.com/DAGroup-PKU/HumanNet) · [数据集文件](https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet/files) · [Hugging Face 万小时子集](https://huggingface.co/datasets/DAGroup-PKU/HumanNet)

## 数据集介绍

**HumanNet** 是面向细粒度行为理解、人体运动建模、人–物交互与具身智能预训练的大规模人类中心视频语料。技术报告介绍的整体语料规模约为 **100 万小时**，同时涵盖第一人称与第三人称视频，并围绕语言描述、运动信息、手部与身体信号等组织标注。

本次开放的 **HumanNet Ego** 是从上述百万小时 HumanNet 整体语料中发布的约 **20 万小时第一视角子集**，记录人类在真实环境中的活动与操作经验，可用于视频学习、数据筛选、动作标注及具身智能等研究。

## 发布范围与数据获取

> **本 ModelScope 仓库提供完整的约 20 万小时 HumanNet Ego 数据。由于存储空间限制，[Hugging Face 仓库](https://huggingface.co/datasets/DAGroup-PKU/HumanNet)仅提供其中约 1 万小时的子集；需要完整 Ego 版本，请使用本 ModelScope 仓库。**

| 数据范围 | 规模与内容 | 获取方式 |
| --- | --- | --- |
| **HumanNet 整体语料** | 约 **100 万小时**，涵盖第一人称与第三人称视频 | 整体研究见 [HumanNet 论文](https://arxiv.org/abs/2605.06747) |
| **HumanNet Ego 发布子集** | 约 **20 万小时**，来自整体语料的第一视角部分 | 本仓库提供完整 Ego 版本：[查看数据集文件](https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet/files) |
| **Hugging Face 子集** | 约 **1 万小时**，是上述 Ego 发布子集的一部分 | [Hugging Face 下载入口](https://huggingface.co/datasets/DAGroup-PKU/HumanNet) |

### 下载完整 20 万小时 Ego 数据

请在本页面的[「数据集文件」](https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet/files)中查看数据，并使用页面提供的「下载数据集」入口获取下载指引。

本仓库采用申请与审核的访问方式。请登录 ModelScope，阅读适用的使用条款并申请访问；审核通过后，使用获得权限的账号下载。

安装 Git LFS 后，也可通过 Git 下载，并使用具有访问权限的 ModelScope 账号完成认证：

~~~bash
git lfs install
git clone https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet.git
~~~

### 获取 Hugging Face 万小时子集

如果只需较小规模的数据，可访问 [Hugging Face 上的 HumanNet](https://huggingface.co/datasets/DAGroup-PKU/HumanNet)。该仓库因存储空间限制，仅提供约 **1 万小时**数据；完整约 **20 万小时 Ego 版本**请从 ModelScope 获取。

## HumanNet 演示视频

<video style="max-width:100%;height:auto" controls preload="metadata" src="https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet/resolve/master/resource/humannet-demo.mp4" width="640"></video>

[打开或下载 HumanNet 演示视频](https://www.modelscope.cn/datasets/DAGroup-PKU/HumanNet/resolve/master/resource/humannet-demo.mp4)

## 引用

如果在研究中使用 HumanNet，请引用：

~~~bibtex
@misc{deng2026humannetscalinghumancentricvideo,
  title={HumanNet: Scaling Human-centric Video Learning to One Million Hours},
  author={Yufan Deng and Daquan Zhou},
  year={2026},
  eprint={2605.06747},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2605.06747}
}
~~~
