ASL-Live-Video-Translation/how2sign-realigned-mediapipe

How2Sign re-aligned MediaPipe landmarks

MediaPipe pose and hand landmarks for a subset of How2Sign, cut at the manually re-aligned sentence timestamps. Built for an ASL-to-English translation course project (CS6170 capstone); the code that produces every file is at github.com/kilpy/ASL-Live-Video-Translation.

How it was made

1. Re-aligned English translations (how2sign_realigned_{train,val,test}.csv) from the aipieces/How2Sign mirror, pinned to revision 1231830fc1e8d77555a245ca22353b144e168111.
2. A 1,000-word vocabulary built from the train split only; we keep sentences whose words are all in it.
3. Segments cut from the full frontal green-screen videos at START_REALIGNED–END_REALIGNED, not from the pre-cut clips, whose boundaries follow the original speech.
4. MediaPipe Holistic run at 15 fps (matching SHuBERT). Missing frames are filled by interpolation.

Contents

Path	What it holds
splits/vocab.txt	The vocabulary, most frequent first
splits/{train,val,test}.txt	Sentence IDs (SENTENCE_NAME) in each split
splits/smoke_{train,val}.txt	Small subsets for a quick CPU test run
video_index.json	Which mirror shard holds each full video
landmarks/{split}/{SENTENCE_NAME}.npz	Per-frame landmarks, float16
smoke/	Landmarks for the smoke subsets only

Each frame keeps 7 upper-body points (nose, shoulders, elbows, wrists) and 21 points per hand, x and y only. No video frames and no face landmarks are included.

Loading

Always pin a revision so results are reproducible:

from huggingface_hub import snapshot_download

path = snapshot_download(
    "ASL-Live-Video-Translation/how2sign-realigned-mediapipe",
    repo_type="dataset",
    revision="<commit hash>",
    allow_patterns=["smoke/*", "splits/*"],
)

Splits and known limitations

• How2Sign's official train/val/test splits are kept; no source video appears in more than one split.
• The same signers appear in every split, so scores may overstate performance on new signers.
• Train sentences come from the videos with the most in-vocabulary sentences, to limit downloads. This reduces topic and signer variety compared with a random sample.
• The 1,000-word filter keeps under a fifth of How2Sign's sentences and favors shorter, simpler ones.

License

How2Sign is © its authors and released under CC BY-NC 4.0. This derived dataset is shared under the same terms: non-commercial use only, with attribution.

Citation

Please cite the original dataset:

@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 = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2021},
  doi       = {10.1109/CVPR46437.2021.00276}
}