How2Sign-MV3D

Final reconstructed SMPL-X signing-motion sequences produced by the MV-Sign multi-view curation pipeline. This is a derived motion resource, not original 3D motion-capture ground truth.

Project: https://github.com/ronggong000/MV-Sign Paper: https://arxiv.org/abs/2611.07811 This dataset is associated with the MV-Sign research project. Please cite the MV-Sign paper and the original source dataset when using this resource.

Contents

There is one final motion NPZ per completed sample. Frontal and side estimates are jointly checked during curation; the released final sequence uses the frontal body coordinate representation. Side-only intermediate sequences, raw videos, OBJ meshes, model weights, and SMPL-X/MANO model assets are not included.

Split	Sequences
Train	30,903
Validation	1,711
Test	2,325
Total	34,939

These counts describe the regenerated release, not a rerun of the submitted paper's training experiments. From 35,241 paired input clips, 289 were excluded by input screening and 13 failed processing. Another 22 frontal validation clips lacked a corresponding side clip and were not in the paired processing manifest. See processing_audit.json.

Download and Verify

Download the data/ tar files and the metadata from the Files tab, or use:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="mage0001/How2Sign-MV3D", repo_type="dataset", local_dir="How2Sign-MV3D")

cd How2Sign-MV3D
sha256sum -c SHA256SUMS
mkdir -p extracted
for archive in data/*.tar; do tar -xf "$archive" -C extracted; done

The tar files contain train/<sample_id>.npz, val/<sample_id>.npz, or test/<sample_id>.npz. NPZ arrays are already compressed, so the outer tar is intentionally uncompressed. samples.csv records the split, archive, member, frame count, FPS, byte size, and per-NPZ SHA-256. manifest.json records archive counts and checksums. All released NPZ payloads are preserved byte-for-byte from the final frontal outputs.

Array Conventions

Load with numpy.load(path, allow_pickle=False). poses is a per-frame SMPL-X local axis-angle pose in radians (55 joints, 3 components); shape and field metadata are in array_schema.json. betas contains fitted shape parameters. Read frame rates from mocap_framerate and samples.csv; do not assume 30 FPS. Non-sign pruning and synthetic motion blends mean released frame indices are not a direct copy of source-video indices. The arrays source_frame_index, source_padding_mask, and synthetic_bridge_mask preserve the saved temporal bookkeeping.

trans, cam_trans, and saved joint/projection fields have different meanings: do not treat trans alone as a calibrated global camera translation, or assume saved tracking joints were recomputed from the final repaired mesh. This release is not camera-calibrated ground-truth geometry. The companion pipeline documentation is being prepared.

Source, Changes, and Terms

Based on How2Sign, released by its authors under CC BY-NC 4.0. Credit the original authors, retain attribution and license information, and use for noncommercial research. Changes comprise automatic body/hand reconstruction, cross-view synchronization and repair, non-sign pruning, and smoothing. This derived resource is not endorsed by the How2Sign authors. SMPL-X/MANO model files require their own licenses and are not distributed here.

Please cite How2Sign and MV-Sign when using this resource. MV-Sign's final bibliographic entry will be added with the paper release.

@inproceedings{duarte2021how2sign,
  title={How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language},
  author={Duarte, Amanda and Palaskar, Shruti and Ventura, Lucas and Ghadiyaram, Deepti and DeHaan, Kenneth and Metze, Florian and Torres, Jordi and Giro-i-Nieto, Xavier},
  booktitle={CVPR},
  year={2021}
}