3DMPB-dataset

Dataset proposed in "Pose2UV: Single-shot Multi-person Mesh Recovery with Deep UV Prior"
[ProjectPage] [paper] @ TIP2022.

Samples of 3DMPB Dataset

3DMPB is a multi-person dataset in the outdoor sport field with human interaction occlusion and image truncation. This dataset provides annotations including bounding-box, human 2D pose, SMPL model annotations, instance mask and camera parameters.

Statistics

Note: Not all the instances in this dataset are annotated in consideration of some inaccurate annotations or wrong relative occlusion.

	bbox	keypoint	mask	SMPL
Images	13,665	13,665	13,665	13,665
Persons	25,122	25,122	25,122	25,122

Download Links

[Baidu Netdisk] [Google Drive]

We also provide full videos for 3DMPB dataset.

[3DMPB-video]

Visualization

Requirements

• python3
• numpy
• pytorch
• pyrender

Prepare dataset

Please download the dataset and extract under ./3DMPB. Due to the licenses, please download SMPL model file here. The folder structure is shown as follows:

|-- 3DMPB
`-- |-- images
    |   |-- 000000.jpg
    |   |-- 000001.jpg
    |   |-- 000002.jpg
    |   |-- ...
    |-- masks
    |   |-- 000000_00.jpg
    |   |-- 000000_01.jpg
    |   |-- 000000_02.jpg
    |   |-- 000001_00.jpg
    |   |-- 000002_00.jpg
    |   |-- ...
    |-- annot.json
|-- data
`-- |-- SMPL_NEUTRAL.pkl    

Visualization

visualize 2D keypoints and boundingbox

python vis_dataset.py --dataset_dir 3DMPB  --output_dir output 

visualize 3D meshes

python vis_dataset.py --dataset_dir 3DMPB  --output_dir output --vis_smpl True

Citation

If you find this dataset useful for your research, please consider citing the paper.

@article{huang2022pose2uv,
  title={Pose2UV: Single-shot Multi-person Mesh Recovery with Deep UV Prior},
  author={Huang, Buzhen and Zhang, Tianshu and Wang, Yangang},
  journal={IEEE Transactions on Image Processing},
  year={2022},
  volume={31},
  pages={4679-4692}
}