SPHP is collected with a proprietary Novatek Motion Vector Sensor and involves human participants. Access is granted for non-commercial academic research only. You must not redistribute the raw data, and you agree to the SPHP data-use agreement. Requests are reviewed and approved manually by the authors / lab.

SPHP: Sparse and Privacy-enhanced Representation for Human Pose Estimation

BMVC 2023 · Paper · Project Page · Code · Video

Ting-Ying Lin*, Lin-Yung Hsieh*, Fu-En Wang, Wen-Shen Wuen, Min Sun — National Tsing Hua University · Novatek Microelectronics Corp.

Dataset summary

SPHP is an in-house dataset for Human Pose Estimation (HPE) built from a proprietary Motion Vector Sensor (MVS). Instead of RGB video, each frame is represented by a sparse, privacy-enhanced pair of an edge image and a two-directional motion-vector image, together with ground-truth labels for 13 body joints and corresponding grayscale images.

40 participants (20 male, 20 female)
16 fitness-related actions, grouped into 4 classes (C1 upper-body, C2 lower-body, C3 slow whole-body, C4 fast whole-body)
Modalities per frame: GR (grayscale), EDG (edge), MVH/MVV (horizontal/vertical motion vectors), plus pose_change ground-truth joint labels

Files

File	Size	Notes
`Master.tar.gz`	~27 GB	Master-view captures
`Slave.tar.gz`	~28 GB	Slave-view captures (same structure as Master)

After extraction (tar zxvf), the structure is:

data/
├── calibrate.npy
├── Master/
│   └── sXX/                # subject id
│       └── <action_id>/    # 01 .. 16
│           ├── EDG/          # 300 png (edge images)
│           ├── MVH/          # 300 png (horizontal motion vector)
│           ├── MVV/          # 300 png (vertical motion vector)
│           └── pose_change/  # 300 npy (ground-truth joint labels)
└── Slave/                   # same structure as Master

Training/testing subject splits and full usage are documented in the official code repository.

Intended use

Research on efficient / privacy-preserving human pose estimation, sparse representations, and sparse-convolution models.

⚠️ License & access terms

This dataset is not released under a standard open license. It was collected with a proprietary, patent-pending Motion Vector Sensor from Novatek Microelectronics, and the original distribution requires signing a data-use agreement (via the project's Google Form).

Use is limited to non-commercial academic research.
Redistribution of the raw data is not permitted without authorization from the authors and Novatek.
Human-subject data: although the representation is privacy-enhanced (edge / motion vectors, not raw video), please respect the participants' consent terms.

If you need access, please contact the authors / lab and complete the required agreement.

Citation

@inproceedings{lin2023sparse,
  title     = {Sparse and Privacy-enhanced Representation for Human Pose Estimation},
  author    = {Lin, Ting-Ying and Hsieh, Lin-Yung and Wang, Fu-En and Wuen, Wen-Shen and Sun, Min},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2023}
}

Downloads last month: 6

Total file size:

59 GB

Paper for VSLab2026/SPHP

Sparse and Privacy-enhanced Representation for Human Pose Estimation

Paper • 2309.09515 • Published Sep 18, 2023