FreeMan

Jiong WANG*¹, Fengyu Yang*¹, Wenbo Gou¹, Bingliang Li¹, Danqi Yan¹, Ailing Zeng³, Yijun Gao²,
Junle Wang², Ruimao Zhang^{1, †}
The Chinese University of Hong Kong (Shenzhen)¹, Tencent², IDEA³
^*Indicates Equal Contribution; ^†Indicates Corresponding Authors

News

📅 [2023.9] FreeMan is ready on Huggingface & OpenXLab. Start from Download Page!
📅 [2023.9] Technical report released. Please join FreeMan Benchmark Group on Google Group and Subscribe updates.

Overview

11M+ Frames
8000 videos

10 Scenarios
27 Real locations

8 views
40 Subjects

3D Pose Annotation

Dataset	Environment	#Subject	#Action	#Scene	#Sequence	#Frame	#Camera	FPS
HumanEva	Laboratory	4	6	1	168	80K	7	30
CMU Panoptic	Laboratory	8	5	1	65	154M	31	30
MPI-INF-3DHP	Real Scene	8	8	1	16	1.3M	14	30
3DPW	Real Scene	7	47	4	60	51K	1	30
Human3.6M	Laboratory	9	15	1	840	3.6M	4(Fixed)	30
AIST	Laboratory	30	10	1	1408	10.1M	9(Fixed)	30
HuMMan	Laboratory	1000	500	1	400K	60M	11(Fixed)	30
HuMMan-Released	Laboratory	132	20	1	4466	278K	11(Fixed)	30
FreeMan	Real Scene	40	123	10^†	8000	11.3M	8(Movable)	30/60

^†FreeMan includes 10 type of scenes that corresponds to 29 locations. Fixed means cameras are fixed within the whole dataset, while our cameras are movable and camera poses vary among video sequences.

Pose annotation pipeline

More examples for Data and Annotations can be found in corresponding pages.

Paper (Arxiv)

If you find FreeMan useful and used it in your research, please cite our paper.

Download

FreeMan dataset is open-sourced under the license of CC-BY-NC-4.0 and is for Research Purpose ONLY. No commercial usage is allowed.
To download FreeMan data, users are required to complete following steps. Detailed instructions in Download page.
1. Download and sign the dataset usage agreement.
2. Complete the Application Form or 申请表格 for information backup.
3. Submit an access application on data host platform: Huggingface or OpenXLab(Coming).

Data Structure

FreeMan consists of 2 subsets corresponding to 30FPS & 60FPS, respectively. Each subset data includes RGB videos, camera parameters, 2D & 3D human keypoints, SMPL parameters, bounding boxes. File structure is shown as below. Data are grouped into different folders by data types.
yyyymmdd_xxxxxxxx_subjNN is the unique name for each session. yyyymmdd refer to capture date, xxxxxxxx refer to uuid of the session, NN refer to subject id. Details can be found in Documentations.

            
            FreeMan
              ├── 30FPS
              │   ├── bbox2d
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN.npy
              │   │   └── ...
              │   ├── keypoints2d
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN.npy
              │   │   └── ...
              │   ├── keypoints3d
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN.npy
              │   │   └── ...
              │   ├── motions
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN_view0.npy
              │   │   ├── ...
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN_view8.npy
              │   │   └── ...
              │   ├── cameras
              │   │   ├── yyyymmdd_xxxxxxxx_subjNN.json
              │   │   └── ...
              │   └── videos
              │        ├── yyyymmdd_xxxxxxxx_subjNN
              │        │   └── vframes
              │        │       ├── c01.mp4
              │        │       ├── ...
              │        │       └── c08.mp4
              │        ├── ...
              ├── 60FPS
              │   ├── ...

License & Statement of Responsibility

FreeMan dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

All actors involved in the work are recuited on basis of voluntary and well informed of data collection and usage. Personal information of actors are disclosed.

Team Members

Jiong Wang, Fengyu Yang
Co-First authors. This work is done when Jiong is an MPhil student and Fengyu is an Msc student supervised by Prof. Ruimao Zhang at The Chinese University of Hong Kong (Shenzhen).

Wenbo Gou, Danqi Yan, Bingliang Li
This work is done when they are research assistant at the research group supervised by Prof. Ruimao Zhang. Bingliang is also Msc student at School of Data Science, The Chinese University of Hong Kong, Shenzhen.

Ailing Zeng
Dr. Zeng is affiliated with IDEA and she serves as a technical advisor in this project.

Yijun Gao, Junle Wang
Mr. Gao and Dr. Wang are affiliated with TEG, Tencent. Yijun provides support on data collection tools and Junle is the leader of team in Tencent. Thanks for their support.

Ruimao Zhang
Prof. Zhang is the corresponding author of this work. He's also the supervisor of students participating in this work.

Contact Us

E-mail: jiongwang@link.cuhk.edu.cn, zhangruimao@cuhk.edu.cn
Address: Daoyuan Building, The Chinese Univeristy of Hong Kong, Shenzhen

FreeMan: Towards benchmarking 3D HPE in the wild