MVSign

MVSign is a multi-view Chinese Sign Language dataset for photorealistic and drivable 3D sign avatar modeling. The dataset is introduced in PHOSA: Photorealistic 3D Sign Avatar Modeling and Benchmark (ECCV 2026).

MVSign was co-designed with Deaf experts and collected under IRB approval. It captures fluent native Chinese Sign Language signers with synchronized multi-view RGB cameras and provides annotations for avatar reconstruction, rendering, and animation, including camera calibration, body-part segmentation, 3D keypoints and SMPL-X parameters.

Highlights

• 5 native Chinese Sign Language signers: 3 female and 2 male signers.
• 16 synchronized RGB cameras at 2048 x 2448 resolution and 25 FPS.
• About 23K temporal frames per signer, about 115K temporal frames in total.
• Dedicated camera layout for both full-body coverage and fine-grained head/hand capture.
• Sign content covering 109 basic hand shapes and daily-use sign sentences.
• Rich annotations for sign avatar modeling: segmentations, 3D keypoints, SMPL-X parameters and camera calibration.

Dataset Structure

The top-level structure is:

MVSign/
|-- README.md
|-- female1/
|-- female2/
|-- female3/
|-- male1/
|-- male2/
|-- scripts/

Each subject directory follows the same structure:

<subject>/
|-- calibration.csv
|-- keypoints3d.npy
|-- smplx_params.npz
|-- valid_frames.npy
`-- data/
    `-- <sequence_id>/
        |-- <sequence_id>.tar.000.part
        |-- <sequence_id>.tar.001.part
        |-- ...
        `-- <sequence_id>.tar.NNN.part

The files under data/<sequence_id>/ are split tar archives. After concatenating and extracting a sequence archive, the extracted directory contains the image data and segmentation annotations:

<subject>/<sequence_id>/
|-- images/
|   `-- <camera_name>/
|       `-- <frame_id>.jpeg
`-- segmentations/
    `-- <camera_name>/
        `-- <frame_id>.png

Annotation Files

calibration.csv stores camera parameters for each subject. The provided script scripts/read_camera_params.py converts the CSV fields into camera intrinsics and extrinsics. The rotation vector fields rx, ry, and rz are converted to a rotation matrix with Rodrigues transformation; tx, ty, and tz form the translation vector. The normalized intrinsic fields fx, fy, px, and py are scaled by image width and height.

keypoints3d.npy stores per-frame 3D skeleton keypoints obtained from multi-view geometric triangulation and optimization.

smplx_params.npz stores the fitted SMPL-X parameter sequence for the subject, including body, hand, and facial parameters used for sign avatar modeling.

valid_frames.npy stores the usable frame indices selected by the Motion-aware Data Sampling Strategy. This strategy filters motion-blurred frames and balances the distribution of sign gesture types.

segmentations/ stores body-part segmentation maps. The helper script scripts/read_mask.py can extract background, hand, and head masks from the segmentation color labels.

Download and Extraction

Clone the dataset with Git LFS or download it with the Hugging Face CLI:

git lfs install
git clone https://huggingface.co/datasets/naaaaapi/MVSign

or:

huggingface-cli download naaaaapi/MVSign --repo-type dataset --local-dir MVSign

Each sequence is stored as split tar parts. Concatenate all parts in order and extract the tar archive:

cd MVSign

sequence_dir=female1/data/21238139
sequence_id=21238139

cat ${sequence_dir}/${sequence_id}.tar.*.part > ${sequence_dir}/${sequence_id}.tar
tar -xf ${sequence_dir}/${sequence_id}.tar -C ${sequence_dir}

The part indices are zero-padded, so the shell glob order is the correct archive order.

Ethics and License

This dataset was collected under IRB approval. All participants provided informed written consent for public release of anonymized data.

MVSign is released under the CC BY-NC 4.0 license. It is intended for non-commercial research use. Users should follow the license terms and use the data responsibly, especially when working with human subjects and sign language data.

Citation

If you use MVSign or PHOSA in your research, please cite:

@inproceedings{wang2026phosa,
  title  = {PHOSA: Photorealistic 3D Sign Avatar Modeling and Benchmark},
  author = {Wang, Haodong and Hu, Hezhen and Zhou, Wengang and Li, Houqiang},
  booktitle = {ECCV},
  year   = {2026}
}