Motion-X: A Large-scale 3D Expressive Whole-body Human Motion Dataset

This repository contains the implementation of the following paper:

Motion-X: A Large-scale 3D Expressive Whole-body Human Motion Dataset
Jing Lin^😎12, Ailing Zeng^😎1, Shunlin Lu^😎13, Yuanhao Cai², Ruimao Zhang³, Haoqian Wang², Lei Zhang^1
😎Equal contribution. ¹International Digital Economy Academy ² Tsinghua University ³The Chinese University of Hong Kong, Shenzhen

🥳 News

• [2023.12.22] We update the sequential motion text descriptions (text_v1.1) augmented by Vicuna 1.5 to enhance the standardization and diversity of text. Please download via this link and replace it with the original file motionx_seq_text.zip. Many thanks to Linghao Chen for polishing the text labels!
• [2023.11.15] We release the rendered SMPL-X visualization of all subsets on DDS platform for quick content viewing.
• [2023.11.15] We release the HumanTOMATO motion representation (tomato representation) and split files.
• [2023.10.26] We release the Motion-X-V1, which provides semantic text labels corresponding to SMPL-X sequences, facial expression motions, and the corresponding texts for augmenting some motions without facial expressions. Please kindly check your email!
• [2023.10.26] We release a high-quality monocular dataset named IDEA400 as a subset of Motion-X, which contains rich expressions and gestures. See this video for more details.

📜 TODO

• Release whole-body pose descriptions.
• Gathering more motion datasets (e.g., music-to-dance, audio-to-gesture motions).
• Release Videos after the agreement of video owners.
• Release audio and music if motions are needed.

Stay tuned!

🥳 Highlight Motion Samples

📊 Table of Contents

1. General Description
2. Dataset Download
3. Experiments
4. Citing

📜 General Description

We propose a high-accuracy and efficient annotation pipeline for whole-body motions and the corresponding text labels. Based on it, we build a large-scale 3D expressive whole-body human motion dataset from massive online videos and eight existing motion datasets. We unify them into the same formats, providing whole-body motion (i.e., SMPL-X) and corresponding text labels.

Labels from Motion-X:

• Motion label: including 15.6M whole-body poses and 81.1K motion clips annotation, represented as SMPL-X parameters.
• Text label: (1) 15.6M frame-level whole-body pose description and (2) 81.1K sequence-level semantic labels.
• Other modalities: RGB videos, audio, and music information.

Supported Tasks:

• Text-driven 3d whole-body human motion generation
• 3D whole-body human mesh recovery
• Others: Motion pretraining, multi-modality pre-trained models for motion understanding and generation, etc.

Dataset	Clip Number	Frame Number	Website	License	Downloading Link
AMASS	26K	5.4M	AMASS Website	AMASS License	AMASS Data
EgoBody	1.0K	0.4M	EgoBody Website	EgoBody License	EgoBody Data
GRAB	1.3K	0.4M	GRAB Website	GRAB License	GRAB Data
IDEA400	12.5K	2.6M	IDEA400 Website	IDEA400 License	IDEA400 Data
AIST++	1.4K	0.3M	AIST++ Website	AIST++ License	AIST++ Data
HAA500	5.2K	0.3M	HAA500 Website	HAA500 License	HAA500 Data
HuMMan	0.7K	0.1M	HuMMan Website	HuMMan License	HuMMan Data
BAUM	1.4K	0.2M	BAUM Website	BAUM License	BAUM Data
Online Videos	32.5K	6.0M	---	---	Online Data
Motion-X (Ours)	81.1K	15.6M	Motion-X Website	Motion-X License	Motion-X Data

📥 Dataset Download

We disseminate Motion-X in a manner that aligns with the original data sources. Here are the instructions:

1. Request Authorization

Please fill out this form to request authorization to use Motion-X for non-commercial purposes. Then you will receive an email and please download the motion and text labels from the provided downloading links. （Note: We updated the Baidu Disk link of motionx_seq_face_text.zip and motionx_face_motion.zip on October 29, 2023. Thus, if you download these zips via Baidu Disk before October 29, please fill out the form and download again.）

Please collect them as the following directory structure:

../datasets  

├──  motion_data
  ├── smplx_322
    ├── idea400
    ├── ...
├──  face_motion_data
  ├── smplx_322
    ├── humanml
    ├── EgoBody
    ├── GRAB
├── texts
  ├──  semantic_labels
    ├── idea400
    ├── ...
  ├──  face_texts
    ├── humanml
    ├── EgoBody
    ├── GRAB
    ├── idea400
    ├── ...

2. Non-Mocap Subsets

For the non-mocap subsets, please refer to this link for a detailed instruction, notably:

• We do not distribute the original RGB videos. We provide the motion and text labels annotated by our team.
• Due to license and quality consideration, we do not provide NTU-RGBD120. Instead, we build IDEA400, which includes 400 daily actions (covering NTU-RGBD120). Please refer to this video for a detailed introduction.

3. Mocap Subsets

For the mocap datasets (i.e., AMASS, GRAB, EgoBody), please refer to this link for a detailed instruction, notably:

• We do not distribute the original motion data.
• We only provide the text labels and facial expressions annotated by our team.

The AMASS and GRAB datasets are released for academic research under custom licenses by Max Planck Institute for Intelligent Systems. To download AMASS and GRAB, you must register as a user on the dataset websites and agree to the terms and conditions of each license:

https://amass.is.tue.mpg.de/license.html

https://grab.is.tuebingen.mpg.de/license.html

Finally, the datasets folder is collected as the following directory structure:

../datasets  

├──  motion_data
  ├── smplx_322
    ├── humanml
    ├── EgoBody
    ├── GRAB
    ├── idea400
    ├── ...
├── texts
  ├──  semantic_labels
    ├── idea400
    ├── ...
  ├──  face_texts
    ├── humanml
    ├── EgoBody
    ├── GRAB
    ├── idea400
    ├── ...

🚀 Data Loading

• To load the motion and text labels you can simply do:

  import numpy as np
  import torch
  
  # read motion and save as smplx representation
  motion = np.load('motion_data/smplx_322/000001.npy')
  motion = torch.tensor(motion).float()
  motion_parms = {
              'root_orient': motion[:, :3],  # controls the global root orientation
              'pose_body': motion[:, 3:3+63],  # controls the body
              'pose_hand': motion[:, 66:66+90],  # controls the finger articulation
              'pose_jaw': motion[:, 66+90:66+93],  # controls the yaw pose
              'face_expr': motion[:, 159:159+50],  # controls the face expression
              'face_shape': motion[:, 209:209+100],  # controls the face shape
              'trans': motion[:, 309:309+3],  # controls the global body position
              'betas': motion[:, 312:],  # controls the body shape. Body shape is static
          }
  
  # read text labels
  semantic_text = np.loadtxt('semantic_labels/000001.npy')     # semantic labels 

💻 Experiments

Validation of the motion annotation pipeline

Our annotation pipeline significantly surpasses existing SOTA 2D whole-body models and mesh recovery methods.

Benchmarking Text-driven Whole-body Human Motion Generation

Comparison with HumanML3D on Whole-body Human Motion Generation Task

Impact on 3D Whole-Body Human Mesh Recovery

🤝 Citation

If you find this repository useful for your work, please consider citing it as follows:

@article{lin2023motionx,
  title={Motion-X: A Large-scale 3D Expressive Whole-body Human Motion Dataset},
  author={Lin, Jing and Zeng, Ailing and Lu, Shunlin and Cai, Yuanhao and Zhang, Ruimao and Wang, Haoqian and Zhang, Lei},
  journal={Advances in Neural Information Processing Systems},
  year={2023}
}