Dataset Viewer
Auto-converted to Parquet Duplicate
sentence_id
stringlengths
25
44
points
listlengths
1
1.09k
detected
listlengths
1
1.09k
signer_id
stringclasses
8 values
-1XUnputrgk_0-5-rgb_front
[ [ [ 0.57568359375, 0.343505859375 ], [ 0.67724609375, 0.53857421875 ], [ 0.50830078125, 0.484619140625 ], [ 0.71826171875, 0.79248046875 ], [ 0.43701171875, 0.73779296875 ], [ 0.64404296875, 0.96875 ...
[ [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, ...
5
-1XUnputrgk_1-5-rgb_front
[ [ [ 0.56591796875, 0.34716796875 ], [ 0.669921875, 0.52197265625 ], [ 0.498046875, 0.50830078125 ], [ 0.7177734375, 0.75537109375 ], [ 0.44775390625, 0.74853515625 ], [ 0.65966796875, 0.72509765625 ...
[ [ true, true, true ], [ true, false, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, false, true ], [ true...
5
-1XUnputrgk_10-5-rgb_front
[ [ [ 0.56787109375, 0.3466796875 ], [ 0.6708984375, 0.5400390625 ], [ 0.48681640625, 0.54150390625 ], [ 0.71630859375, 0.74072265625 ], [ 0.437255859375, 0.73681640625 ], [ 0.77978515625, 0.83056640625...
[ [ true, false, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true, true, true ], [ true,...
5
-1XUnputrgk_14-5-rgb_front
[[[0.55712890625,0.349853515625],[0.66650390625,0.5166015625],[0.497802734375,0.517578125],[0.698242(...TRUNCATED)
[[true,true,false],[true,false,false],[true,true,true],[true,false,true],[true,true,true],[true,true(...TRUNCATED)
5
-1XUnputrgk_17-5-rgb_front
[[[0.5556640625,0.357421875],[0.66259765625,0.51953125],[0.489501953125,0.52099609375],[0.7133789062(...TRUNCATED)
[[true,false,true],[true,true,false],[true,true,true],[true,true,true],[true,true,true],[true,true,t(...TRUNCATED)
5
-1XUnputrgk_18-5-rgb_front
[[[0.556640625,0.3740234375],[0.66796875,0.5390625],[0.4931640625,0.53173828125],[0.720703125,0.7714(...TRUNCATED)
[[true,false,true],[true,false,true],[true,false,true],[true,false,true],[true,false,true],[true,fal(...TRUNCATED)
5
-1XUnputrgk_22-5-rgb_front
[[[0.5703125,0.356689453125],[0.68505859375,0.544921875],[0.5078125,0.5185546875],[0.72802734375,0.8(...TRUNCATED)
[[true,false,true],[true,false,true],[true,false,true],[true,false,true],[true,false,true],[true,fal(...TRUNCATED)
5
-1XUnputrgk_23-5-rgb_front
[[[0.5576171875,0.3466796875],[0.6708984375,0.5234375],[0.50048828125,0.51953125],[0.734375,0.768066(...TRUNCATED)
[[true,true,true],[true,true,true],[true,true,true],[true,true,true],[true,true,true],[true,true,tru(...TRUNCATED)
5
-1XUnputrgk_5-5-rgb_front
[[[0.56103515625,0.36474609375],[0.67431640625,0.52197265625],[0.5029296875,0.52392578125],[0.745117(...TRUNCATED)
[[true,true,true],[true,true,true],[true,true,false],[true,true,true],[true,true,true],[true,true,tr(...TRUNCATED)
5
-8tIt1vEinQ_1-8-rgb_front
[[[0.541015625,0.357177734375],[0.64892578125,0.55712890625],[0.46875,0.546875],[0.65625,0.811035156(...TRUNCATED)
[[true,false,false],[true,false,false],[true,false,true],[true,false,false],[true,false,true],[true,(...TRUNCATED)
8
End of preview. Expand in Data Studio

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
signers.csv Published sentence IDs, splits, video names, and derived signer IDs
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 pose features with Datasets

The Parquet configurations support load_dataset directly:

from datasets import load_dataset

dataset = load_dataset(
    "ASL-Live-Video-Translation/how2sign-realigned-mediapipe",
    "rtmw",  # or "mediapipe", "vitpose", or "dwpose"
)

train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

Choose the configuration mediapipe, vitpose, rtmw, or dwpose. MediaPipe has train and validation data; the other configurations also have test data. Pass split="train", split="validation", or split="test" to load one split. Each row contains sentence_id, signer_id, points (frames × landmarks × 2), and detected; the non-MediaPipe configurations also contain confidence. The arrays retain the NPZ float16 and Boolean values. MediaPipe has 49 points and three detection flags per frame; the other sources have 65 points and per-point detection flags. The existing NPZ files and metadata remain available. Rows contain sentence IDs and pose features; they do not include translation text.

signer_id is the number immediately before -rgb_front in the sentence and video names. This is a derived pseudonymous identifier, not a person's name. The filename groups match the per-signer green-screen video counts in How2Sign supplementary Table S2 exactly for validation and test; the released training annotations contain one fewer video than the table. The source paper states signer IDs are provided but does not explicitly document the filename convention, so this mapping is an inference. signers.csv covers the 2,719 distinct sentences with published features, including test sentences absent from the MediaPipe config.

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.
  • Signers overlap across splits; the test set also contains signer 10, who is absent from training.
  • 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.

Additional pose sources

The dataset includes native whole-body keypoints from ViTPose, RTMW, and DWPose alongside the original MediaPipe landmarks. The new sources use 65 points: 23 COCO-WholeBody body/foot points and 21 points per hand; detailed face points are omitted. Each NPZ contains normalized points with shape (frames, 65, 2), per-point detected flags, and per-point confidence. See each source's landmarks/<source>/metadata.json for model and extraction details.

Select a source and partition by path:

Source Train Validation Test
MediaPipe landmarks/train/ landmarks/val/ Not published
ViTPose landmarks/vitpose/train/ landmarks/vitpose/val/ landmarks/vitpose/test/
RTMW landmarks/rtmw/train/ landmarks/rtmw/val/ landmarks/rtmw/test/
DWPose landmarks/dwpose/train/ landmarks/dwpose/val/ landmarks/dwpose/test/

For example, download only RTMW's training partition with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ASL-Live-Video-Translation/how2sign-realigned-mediapipe",
    repo_type="dataset",
    revision="<dataset commit hash>",
    allow_patterns=["landmarks/rtmw/train/*.npz", "landmarks/rtmw/metadata.json", "splits/*"],
    local_dir="how2sign-features",
)

Use val or test in the pattern to select another partition. The split ID lists contain 10 train and one validation sentence with no feature file because their annotated start times are past the end of their source videos. Those IDs are absent from MediaPipe and all three new sources.

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}
}

Sports and Fitness features

The sports_* configurations use at least 50 decoded source frames and 10 words; every other source frame was processed. source_fps and fps are per-sentence rates. Labels come from the How2Sign research topic tables; conflicting and unknown labels were excluded. Signer IDs are inferred from source filenames. sports/metadata.csv contains sentence and signer metadata; sports/features/ contains source-specific NPZ archives.

Load the Sports and Fitness data by pose estimator and split:

from datasets import load_dataset

train = load_dataset(
    "ASL-Live-Video-Translation/how2sign-realigned-mediapipe",
    "sports_mediapipe",  # also sports_vitpose, sports_rtmw, sports_dwpose
    split="train",  # also validation or test
)
Downloads last month
258