SignBible-MediaPipe / README.md
Haloroute's picture
Upload README.md with huggingface_hub
f7102b1 verified
|
Raw History Blame Contribute Delete
17.3 kB
metadata
task_categories:
  - translation
  - video-classification
language:
  - ase
  - bqn
  - esl
  - eso
  - eth
  - gse
  - ins
  - lsb
  - mis
  - mzy
  - nsi
  - nsp
  - rsl
  - sqs
  - tza
  - ugn
  - xki
license: cc-by-sa-4.0
datasets:
  - bible-nlp/sign-bibles
tags:
  - sign-language
  - mediapipe
  - holistic
  - keypoints
  - pose
  - sign-bibles
  - bible-nlp
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/shard_*-train.parquet
      - split: val
        path: data/shard_*-val.parquet
      - split: test
        path: data/shard_*-test.parquet

SignBible-MediaPipe

MediaPipe Holistic 3D keypoint dataset extracted from the .pose-mediapipe.pose files of bible-nlp/sign-bibles across 17 sign languages, 21 Bible translation projects, and 1,653 source WebDataset .tar shards (~8,103 estimated samples).

  • Storage format: zstd-compressed Parquet shards (data/shard_XXXXX-{split}.parquet), capped at 25 videos per shard to keep memory bounded on long Bible passages.
  • Array encoding: The 8 keypoint and confidence tensors are stored as raw C-contiguous float32 bytes (Value("binary") in Hugging Face Features / binary in Arrow) instead of nested Arrow lists, avoiding RAM blow-up during serialization and streaming.

Keypoints Extraction, Normalization & Face Mesh Masking

  • Source format: pose-format binary .pose-mediapipe.pose files from bible-nlp/sign-bibles (POSE_LANDMARKS 33, FACE_LANDMARKS 468 or 478, LEFT_HAND_LANDMARKS 21, RIGHT_HAND_LANDMARKS 21; POSE_WORLD_LANDMARKS is omitted). The leading person dimension (1, T, ...) is squeezed to (T, ...).
  • Coordinate normalization: x and y pixel coordinates are normalized to [0, 1] by dividing by the pose header / video dimensions (width, height), while z relative depth is preserved verbatim.
  • Invalid & missing detections: Landmarks that are masked in .pose, non-finite (NaN/Inf), or have conf <= 0 are zeroed out (0.0) in both coordinates and confidence so downstream consumers never encounter NaN values. Mask frames/landmarks using *_conf > 0.
  • 468 vs. 478 Face Landmarks Masking: Estonian Sign Language (eso/the_bible_in_estonian_sign_language) stores 468 face mesh points (omitting the 10 MediaPipe iris landmarks 468:478). To keep a uniform (T, 478, 3) / (T, 478) tensor shape across all languages, face_pts[:, 468:478, :] is zero-padded (0.0) and face_conf[:, 468:478] is masked to 0.0. Check meta["face_landmarks_orig"] (468 vs. 478) to distinguish iris availability.

Splits & Extraction Summary

Split Discovered Source .tar Shards Completed Tasks Failed / Empty Samples Uploaded
train 1153 1148 5 5451
val 171 171 0 848
test 329 328 1 1602

  • All .pose-mediapipe.pose shards are 100% extracted: Across the 1,653 source .tar shards in bible-nlp/sign-bibles, all 1,647 shards containing .pose-mediapipe.pose files (1,148 train, 171 val, 328 test) are fully uploaded into 323 Parquet shards (224 train, 34 val, 65 test; 25 rows/shard max).
  • Why 6 source .tar shards appear under Failed / Empty (5 train, 1 test): None of those 6 archives contain .pose-mediapipe.pose files upstream:
    • ase/chronological_bible_translation_in_american_sign_language_119_introductions_and_passages_expanded_with_more_information/shard_00048-train.tar: 10 KB empty archive (0 files; videos > 2 GB omitted upstream).
    • Five shards in mis/chronological_bible_translation_in_west_bengal_sign_language (shard_00014-train.tar, shard_00017-train.tar, shard_00022-train.tar, shard_00025-train.tar, and shard_00019-test.tar): contain .mp4, .pose-dwpose.npz, .json, and .transcripts.json upstream, with 0 .pose-mediapipe.pose files.
  • Unique keys vs. total Parquet rows: Samples Uploaded reports unique sample keys (5,451 train, 848 val, 1,602 test; 7,901 unique keys total). Because a few upstream subsets (bqn, mis/andhra_pradesh, mis/south_sudanese, xki) package the same sample_id into more than one source .tar shard (149 duplicate keys in train, 1 in test), the Parquet shards hold 5,600 (train), 848 (val), and 1,603 (test) rows (8,051 total rows).

Schema & Columns

Each row corresponds to one sign-language video clip (T = duration frames). The 8 keypoint and confidence columns store raw C-contiguous float32 bytes (4 bytes per float) and are reconstructed with np.frombuffer(sample[col], dtype=np.float32).reshape(T, ...):

Column Arrow Type Decoded Shape / Byte Length Description
pose_pts binary float32[T, 33, 3] (T * 33 * 3 * 4 B) Normalized 3D body pose landmarks (x, y, z).
pose_conf binary float32[T, 33] (T * 33 * 4 B) Per-landmark visibility score (0.0 when missing/masked).
face_pts binary float32[T, 478, 3] (T * 478 * 3 * 4 B) Normalized 3D face mesh landmarks (468:478 zero-padded when source has 468 points).
face_conf binary float32[T, 478] (T * 478 * 4 B) Face landmark confidence (0.0 for missing frames, masked points, and 468:478 when padded).
left_hand_pts binary float32[T, 21, 3] (T * 21 * 3 * 4 B) Normalized 3D left-hand landmarks (x, y, z).
left_hand_conf binary float32[T, 21] (T * 21 * 4 B) Left-hand landmark confidence (0.0 when missing/masked).
right_hand_pts binary float32[T, 21, 3] (T * 21 * 3 * 4 B) Normalized 3D right-hand landmarks (x, y, z).
right_hand_conf binary float32[T, 21] (T * 21 * 4 B) Right-hand landmark confidence (0.0 when missing/masked).
transcript string — Concatenated verse text from .transcripts.json (NFC-normalized, \n-joined when multiple segments exist).
fps float32 scalar Frame rate from .pose header (fallback to sample metadata or 30.0).
duration int32 scalar Total number of frames T.
key string — Unique identifier: <split>/<lang>/<project>/<sample_id>.
metadata string JSON JSON object (ensure_ascii=False) containing original and extraction metadata (fields below).

Metadata Fields

Every row's metadata column is a JSON-encoded string containing the original .json fields, verse-aligned .transcripts.json segments, and source shard provenance:

Field Type Description
language dict Language metadata (ISO639-3, BCP-47, name, nameLocal).
project_name string Human-readable Digital Bible Library (DBL) project title.
copyright list[dict] Attribution entries (name, abbr, url) for the sign language translation.
license string Source license string (e.g., "Creative Commons - Attribution-ShareAlike [CC BY-SA]").
source string Digital Bible Library entry URL.
total_frames int Total frame count T of the video / pose sequence.
fps float Video frame rate.
width int Source video frame width in pixels (used for x normalization).
height int Source video frame height in pixels (used for y normalization).
duration_sec float Clip duration in seconds.
bible-ref string Passage reference (e.g., "LUK 1:1-4", "GEN 9:18-29").
split string Dataset split ("train", "val", or "test").
sample_id string Original sample basename inside the WebDataset .tar shard.
lang_code string Three-letter ISO 639-3 language folder code (e.g., "ase", "eso", "xki").
project_slug string Project folder slug in bible-nlp/sign-bibles.
face_landmarks_orig int Original FACE_LANDMARKS count in the source .pose (478 or 468).
transcripts list[dict] Verse/segment annotations from .transcripts.json (text, start_frame, end_frame, language, license, source, biblenlp-vref, bible-ref).
tar_shard string Source .tar path in bible-nlp/sign-bibles.
tar_key string Pipeline task key <split>/<lang>/<project>/shard_XXXXX-<split>.
tar_num_samples int Number of .pose-mediapipe.pose samples in that source .tar shard.

Example metadata JSON (abbreviated):

{
  "language": {"ISO639-3": "eso", "BCP-47": "eso", "name": "Estonian Sign Language", "nameLocal": "Eesti viipekeel"},
  "project_name": "The Bible in Estonian Sign Language",
  "copyright": [{"name": "Deaf Bible Society", "abbr": "Deaf Bible", "url": "https://www.deafbiblesociety.com"}],
  "license": "Creative Commons - Attribution-ShareAlike [CC BY-SA]",
  "source": "https://app.thedigitalbiblelibrary.org/entry?id=6ad9cf57ab084a3b",
  "total_frames": 1390,
  "fps": 25.0,
  "width": 3840,
  "height": 2160,
  "duration_sec": 55.6,
  "bible-ref": "LUK 1:1-4",
  "split": "test",
  "sample_id": "eso_the_bible_in_estonian_sign_language_luk_1_1-4",
  "lang_code": "eso",
  "project_slug": "the_bible_in_estonian_sign_language",
  "face_landmarks_orig": 468,
  "transcripts": [
    {
      "text": "...",
      "start_frame": 0,
      "end_frame": 1389,
      "bible-ref": "LUK 1:1-4",
      "biblenlp-vref": [1, 2, 3, 4]
    }
  ],
  "tar_shard": "eso/the_bible_in_estonian_sign_language/shard_00000-test.tar",
  "tar_key": "test/eso/the_bible_in_estonian_sign_language/shard_00000-test",
  "tar_num_samples": 5
}

Source Subsets (bible-nlp/sign-bibles)

ISO Language Subset (<lang>/<project>) Face Pts Train Val Test Total Shards Est. Samples
ase American Sign Language ase/chronological_bible_translation_in_american_sign_language_119_introductions_and_passages 478 28 5 9 42 182
ase American Sign Language ase/chronological_bible_translation_in_american_sign_language_119_introductions_and_passages_expanded_with_more_information 478 35 5 9 49 179
bqn Bulgarian Sign Language bqn/chronological_bible_translation_in_bulgarian_sign_language 478 70 10 21 101 505
esl Egyptian Sign Language esl/chronological_bible_translation_in_egyptian_sign_language 478 7 2 3 12 60
eso Estonian Sign Language eso/the_bible_in_estonian_sign_language 468 13 2 3 18 90
eth Ethiopian Sign Language eth/chronological_bible_translation_in_ethiopian_sign_language 478 76 11 21 108 540
gse Ghanaian Sign Language gse/chronological_bible_translation_in_ghanaian_sign_language 478 75 11 21 107 535
ins Indian (Delhi) Sign Language ins/chronological_bible_translation_in_indian_delhi_sign_language 478 14 2 4 20 93
lsb Burundian Sign Language lsb/chronological_bible_translation_in_burundian_sign_language 478 76 11 21 108 540
mis Andhra Pradesh Sign Language mis/chronological_bible_translation_in_andhra_pradesh_sign_language 478 77 11 22 110 550
mis Kerala Sign Language mis/chronological_bible_translation_in_kerala_sign_language 478 77 11 22 110 550
mis South Sudanese Sign Language mis/chronological_bible_translation_in_south_sudanese_sign_language 478 93 14 27 134 625
mis West Bengal Sign Language mis/chronological_bible_translation_in_west_bengal_sign_language 478 18 3 5 26 121
mzy Mozambican Sign Language mzy/chronological_bible_translation_in_mozambican_sign_language 478 35 5 10 50 250
nsi Nigerian Sign Language nsi/chronological_bible_translation_in_nigerian_sign_language 478 76 11 21 108 540
nsp Nepali Sign Language nsp/chronological_bible_translation_in_nepali_sign_language 478 7 1 3 11 55
rsl Russian Sign Language rsl/chronological_bible_translation_in_russian_sign_language 478 52 8 15 75 375
sqs Sri Lankan Sign Language sqs/chronological_bible_translation_in_sri_lankan_sign_language 478 2 1 1 4 13
tza Tanzanian Sign Language tza/chronological_bible_translation_in_tanzanian_sign_language 478 75 11 21 107 535
ugn Ugandan Sign Language ugn/chronological_bible_translation_in_ugandan_sign_language 478 75 11 21 107 535
xki Kenyan Sign Language xki/chronological_bible_translation_in_kenyan_sign_language 478 172 25 49 246 1230
Total 17 languages / 21 projects — 468 / 478 1,153 171 329 1,653 8,103

Usage

import json
import numpy as np
from datasets import load_dataset

ARRAY_SHAPES = {
    "pose_pts": (33, 3),
    "pose_conf": (33,),
    "face_pts": (478, 3),
    "face_conf": (478,),
    "left_hand_pts": (21, 3),
    "left_hand_conf": (21,),
    "right_hand_pts": (21, 3),
    "right_hand_conf": (21,),
}


def decode_keypoints(sample: dict) -> dict[str, np.ndarray]:
    """Decode all 8 binary columns in a sample row to float32 numpy arrays."""
    t = int(sample["duration"])
    return {
        col: np.frombuffer(sample[col], dtype=np.float32).reshape(t, *tail)
        for col, tail in ARRAY_SHAPES.items()
    }


# Stream shards directly without downloading the whole dataset
ds = load_dataset("Haloroute/SignBible-MediaPipe", split="train", streaming=True)
sample = next(iter(ds))
kpts = decode_keypoints(sample)
meta = json.loads(sample["metadata"])

print(sample["key"], kpts["pose_pts"].shape, meta.get("face_landmarks_orig"), meta.get("bible-ref"))

# Mask missing detections (confidence > 0.0; iris 468:478 is 0.0 when face_landmarks_orig == 468)
lh_detected_frames = kpts["left_hand_conf"][:, 0] > 0.0
face_valid_mask = kpts["face_conf"] > 0.0

# Slice keypoints for each verse / transcript segment
for seg in meta.get("transcripts", []):
    s, e = int(seg["start_frame"]), int(seg["end_frame"])
    seg_pose = kpts["pose_pts"][s : e + 1]
    print(seg.get("bible-ref"), seg.get("biblenlp-vref"), seg_pose.shape, seg["text"][:80])

Source Notes & Known Issues

  • Videos larger than 2 GB are omitted upstream in bible-nlp/sign-bibles due to PyArrow ArrowCapacityError (huggingface/datasets#7691); see BibleNLP/sign-bibles-dataset#22. This leaves 1 empty archive (ase/chronological_bible_translation_in_american_sign_language_119_introductions_and_passages_expanded_with_more_information/shard_00048-train.tar, 10 KB, 0 samples).
  • Missing .pose-mediapipe.pose in West Bengal Sign Language (mis): Five upstream shards (mis/chronological_bible_translation_in_west_bengal_sign_language/shard_00014-train.tar, shard_00017-train.tar, shard_00022-train.tar, shard_00025-train.tar, and shard_00019-test.tar) contain .mp4, .pose-dwpose.npz, .json, and .transcripts.json files but omit .pose-mediapipe.pose in bible-nlp/sign-bibles, so they produce 0 MediaPipe records here.
  • Duplicate sample_id across upstream .tar shards: Some upstream subsets (bqn, mis/andhra_pradesh, mis/south_sudanese, xki) include the same sample_id in two different .tar shards (149 keys in train, 1 key in test). Every source .tar shard is extracted without deduplicating across .tar shards, yielding 8,051 Parquet rows (7,901 unique keys).
  • Non-square pixel aspect ratio: Some source videos had non-square pixels, which can horizontally squeeze frames and keypoints computed from them; see BibleNLP/sign-bibles-dataset#45.
  • Gloss annotations: Timestamped ASL gloss annotations for 10 videos (679 annotations) are available in bible-nlp/sign-bibles (osb_glosses).

License & Citation

Derived from bible-nlp/sign-bibles under CC BY-SA 4.0. Please check individual sample metadata (copyright, license, source) for specific attribution requirements and cite:

@misc{leong2026osb,
  title        = {Open Sign Bibles ({OSB}): An Open-Access Multilingual Sign Language Bibles Dataset},
  author       = {Leong, Colin and Nemecek, Joshua and Raju, Kavitha and Mathew, Joel and Asari, Vijayan and Leong, Elisabeth},
  year         = {2026},
  howpublished = {Research Square preprint},
  doi          = {10.21203/rs.3.rs-7918028/v1}
}