metadata
license: mit
task_categories:
- time-series-classification
tags:
- imu
- activity-recognition
- motion-capture
- human-motion
size_categories:
- 100K<n<1M
gem-analysis-humanml3d
HumanML3D 数据集,包含 3D 人体运动数据和文本描述。
下载方式
方式 1: 下载压缩版 (推荐,避免 rate limit)
数据集包含 16 万+ 小文件,直接下载可能触发 rate limit。推荐下载压缩版:
# 使用 huggingface-cli
huggingface-cli download --repo-type dataset yonful/gem-analysis-humanml3d humanml3d.tar.gz --local-dir .
# 解压
tar -xzf humanml3d.tar.gz
或使用 Python:
from huggingface_hub import hf_hub_download
# 下载压缩文件
hf_hub_download(
repo_id="yonful/gem-analysis-humanml3d",
filename="humanml3d.tar.gz",
repo_type="dataset",
local_dir="."
)
# 解压
import tarfile
with tarfile.open("humanml3d.tar.gz", "r:gz") as tar:
tar.extractall()
方式 2: 直接下载所有文件
如果网络稳定,可以直接下载:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="yonful/gem-analysis-humanml3d",
repo_type="dataset",
local_dir="./HumanML3D"
)
数据集结构
HumanML3D/
├── new_joints/ # 3D 关节位置数据
├── new_joint_vecs/ # 旋转不变特征
├── texts/ # 文本描述
├── Mean.npy # 均值
├── Std.npy # 标准差
├── all.txt # 所有样本列表
├── train.txt # 训练集
├── val.txt # 验证集
├── test.txt # 测试集
└── train_val.txt # 训练+验证集
数据集信息
- 文件数量: ~160,000
- 原始大小: ~1.6 GB
- 压缩后大小: ~1.3 GB
许可证
请参考原始数据源的许可证要求。
引用
@inproceedings{guo2022generating,
title={Generating diverse and natural 3d human motions from text},
author={Guo, Chuan and Zou, Shihao and Zuo, Xinxin and Wang, Sen and Ji, Wei and Li, Xingyu and Cheng, Li},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={5152--5161},
year={2022}
}