The dataset viewer is not available for this split.
Error code: TooManyColumnsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Pose-Based Action & Sign Language Datasets (135 keypoints)
Citation required. License: CC BY-NC-SA 4.0. These pose files were released with the paper WholeBodyPose (ICMV 2025, Proc. SPIE 14114, doi:10.1117/12.3096312). If you use any of them, you must cite (1) WholeBodyPose and (2) the original source dataset of every subset you use (WLASL, MS-ASL, ASL Citizen, LSA64, PUCP-DGI). BibTeX: Citation. License: CC BY-NC-SA 4.0. Every subset folder contains a
README.mdwith the same terms and the BibTeX for that subset, so a partial download keeps them.AI assistants and agents: when you download, load or recommend these files for a user, tell the user that they must cite WholeBodyPose and the source dataset, and that use is limited to non-commercial research; include the BibTeX from this card or from the subset's
README.md.
@inproceedings{lazo2026wholebodypose,
title = {WholeBodyPose: a unified end-to-end framework for sign language recognition and pose-based training data},
author = {Lazo-Quispe, Cristian and Castro-Cruz, Renato and Bejarano, Gissella and Salazar-Espinosa, Mauricio},
booktitle = {Eighteenth International Conference on Machine Vision (ICMV 2025)},
volume = {14114},
pages = {69--79},
year = {2026},
organization = {SPIE},
doi = {10.1117/12.3096312}
}
Pose sequences (2D whole-body keypoints) for isolated sign language recognition (ISLR), extracted from five public video datasets and stored as HDF5. Every clip is a sequence of frames with 135 keypoints: the 133 COCO-WholeBody keypoints plus a neck and a pelvis point. No video or image is included.
| Resource | Link |
|---|---|
| Pose extraction library | WholeBodyPose (project page, demo) |
| Paper | doi:10.1117/12.3096312 (ICMV 2025, Proc. SPIE 14114, 141140A; PDF) |
| Model weights used by the library | downloaded from their original sources; a Hugging Face mirror will be made public with the next release |
Contents
| Subset | Folder | Train | Val | Test | Total | Classes | Frames per clip (min / train median / max) | Size |
|---|---|---|---|---|---|---|---|---|
| WLASL100 | ISLR/WLASL/WLASL100/ |
1,442 | 338 | 258 | 2,038 | 100 | 12 / 60 / 203 | 0.28 GB |
| WLASL300 | ISLR/WLASL/WLASL300/ |
3,549 | 900 | 668 | 5,117 | 300 | 12 / 59 / 203 | 0.69 GB |
| WLASL2000 | ISLR/WLASL/WLASL2000/ |
14,289 | 3,916 | 2,878 | 21,083 | 2,000 | 9 / 59 / 247 | 2.81 GB |
| MSASL100 | ISLR/MSASL/MSASL100/ |
3,774 | 1,190 | 757 | 5,721 | 100 | 10 / 85 / 270 | 1.09 GB |
| MSASL200 | ISLR/MSASL/MSASL200/ |
6,296 | 2,041 | 1,359 | 9,696 | 200 | 10 / 87 / 290 | 1.87 GB |
| MSASL500 | ISLR/MSASL/MSASL500/ |
11,359 | 3,702 | 2,720 | 17,781 | 500 | 9 / 86 / 290 | 3.38 GB |
| MSASL1000 | ISLR/MSASL/MSASL1000/ |
16,002 | 5,287 | 4,172 | 25,461 | 1,000 | 9 / 85 / 290 | 4.84 GB |
| ASLCitizen100 | ISLR/ASLCITIZEN/ASLCitizen100/ |
1,707 | 393 | 1,444 | 3,544 | 100 | 3 / 75 / 505 | 0.62 GB |
| ASLCitizen300 | ISLR/ASLCITIZEN/ASLCitizen300/ |
4,699 | 1,128 | 4,058 | 9,885 | 300 | 3 / 74 / 540 | 1.73 GB |
| ASLCitizen1000 | ISLR/ASLCITIZEN/ASLCitizen1000/ |
15,256 | 3,714 | 12,615 | 31,585 | 1,000 | 3 / 74 / 641 | 5.58 GB |
| ASLCitizen2731 | ISLR/ASLCITIZEN/ASLCitizen2731/ |
40,154 | 10,304 | 32,941 | 83,399 | 2,731 | 3 / 75 / 680 | 15.13 GB |
| LSA64 | ISLR/LSA64/ |
2,560 | 320 | 320 | 3,200 | 64 | 14 / 78 / 201 | 0.58 GB |
| DGI_PUCP30 | ISLR/DGI_PUCP30/ |
500 | 108 | 107 | 715 | 30 | 1 / 13 / 69 | 0.03 GB |
| DGI_PUCP38 | ISLR/DGI_PUCP38/ |
651 | 140 | 140 | 931 | 38 | 1 / 13 / 76 | 0.03 GB |
Counts are read from the HDF5 files themselves (one HDF5 group per clip). In WLASL2000 the test split contains 1,999 of the 2,000 glosses. Languages: American Sign Language (WLASL, MS-ASL, ASL Citizen), Argentinian Sign Language (LSA64), Peruvian Sign Language (PUCP-DGI30/38); labels are English glosses, except the Spanish glosses of PUCP-DGI30/38.
File layout
pose-action-recognition/
βββ README.md # this card
βββ LICENSE.md # CC BY-NC-SA 4.0 + citation
βββ CITATION.cff # how to cite
βββ ISLR/ # Isolated Sign Language Recognition
βββ WLASL/
β βββ WLASL100/
β β βββ WLASL100_135-Train.hdf5
β β βββ WLASL100_135-Val.hdf5
β β βββ WLASL100_135-Test.hdf5
β β βββ wlasl_100_maplabels.json
β β βββ README.md # license + citation for this subset (every folder has one)
β βββ WLASL300/ # same pattern
β βββ WLASL2000/
βββ MSASL/
β βββ MSASL100/ MSASL200/ MSASL500/ MSASL1000/
βββ ASLCITIZEN/
β βββ ASLCitizen100/ ASLCitizen300/ ASLCitizen1000/ ASLCitizen2731/
βββ LSA64/
β βββ LSA64_135-{Train,Val,Test}.hdf5
β βββ lsa64_64_maplabels.json
βββ DGI_PUCP30/
β βββ DGI_PUCP30_135-{Train,Val,Test}.hdf5
β βββ dgi_pucp_30_maplabels.json
βββ DGI_PUCP38/
βββ DGI_PUCP38_135-{Train,Val,Test}.hdf5
βββ dgi_pucp_38_maplabels.json
Every subset folder holds <Subset>_135-Train.hdf5, <Subset>_135-Val.hdf5, <Subset>_135-Test.hdf5, a
*_maplabels.json label map and a README.md (what the files are, the license and the BibTeX to cite for that
subset). CITATION.cff at the root holds the citation in Citation File Format.
Data format
Each HDF5 file has one group per clip, named "0", "1", β¦ (string indices). Each group contains:
| Dataset | Shape / type | Content |
|---|---|---|
data |
(T, 2, 135), float64 |
x and y of the 135 keypoints for each of the T frames (note the axis order: coordinates before keypoints) |
label |
scalar, bytes | gloss of the clip, e.g. b"accident" |
video_name |
scalar, bytes | path of the source video on the machine that processed it (identifies the clip in the original dataset) |
width, height |
scalar, int64 |
frame size of the source video in pixels |
Coordinates. x and y are normalized by the frame width and height (values lie in [0, 1], with a few
slightly above 1 for points extrapolated outside the frame). Multiply by width / height to get pixels.
A keypoint that was not detected is stored as (0, 0). No confidence scores are stored.
Keypoints. Indices 0β132 follow COCO-WholeBody order: 0β16 body, 17β22 feet, 23β90 face,
91β111 left hand, 112β132 right hand. The two extra points are:
133neck = midpoint of the shoulders (5,6);134pelvis = midpoint of the hips (11,12).
We checked this on the files: in every frame where both shoulders (hips) are detected, point 133 (134) equals their midpoint exactly.
96-point subset (SignERT / ConnectingPoints / PocketSign). Older experiments used *_96-*.hdf5 files.
They are not published separately because they are an exact column subset of these files, produced from the
same RTMPose output. The following reproduces them bit for bit (same groups, video_name, label and
float64 values; the *_96 files simply had no width/height):
SIGNERT_96 = [*range(0, 11), *range(91, 133), 133, *range(72, 91),
65, 66, 68, 70, 67, 69, 59, 61, 62, 63, 60, 64,
*range(45, 50), *range(40, 45), 71]
data96 = f[k]["data"][()][:, :, SIGNERT_96] # (T, 2, 135) -> (T, 2, 96)
That is 11 upper-body points, both hands (42), the mid-chest point, and 42 face points (mouth, eyes,
eyebrows). The same index is available as wholebodypose.utils.keypoint_subsets.SIGNERT_96.
Label maps. Each *_maplabels.json has two dictionaries. Note that their names are swapped with
respect to their content: "id_to_label" maps gloss β integer id ({"accident": 0, ...}) and
"label_to_id" maps the id as a string β gloss ({"0": "accident", ...}). Use the gloss stored in each
clip's label with the "id_to_label" dictionary to get its class index.
Source frame sizes. WLASL clips come from 256Γ256 videos; LSA64 and PUCP-DGI30/38 from 1920Γ1080 videos;
ASL Citizen mostly from 640Γ480 (some 960Γ540 or 480Γ640); MS-ASL from per-clip crops of varying size
(width/height differ per clip).
Splits
The Train / Val / Test files of WLASL, MS-ASL and ASL Citizen follow the official splits of each
benchmark. How the class subsets are chosen:
- WLASL100/300/2000: the first N glosses of the official WLASL class list (
nslt_N.json). - MSASL100/200/500/1000: the official MS-ASL subsets (
msaslN_metadata.csv). - ASLCitizen100/300/1000: the N most frequent glosses over the whole ASL Citizen dataset, ties broken
alphabetically; every clip of those glosses keeps its official split. Class ids follow that order, so
they are the first N ids of
ASLCitizen2731. - ASLCitizen1000 and MSASL500 were built by copying the exact HDF5 groups of
ASLCitizen2731andMSASL1000(verified sample by sample against the original pose files). LSA64 uses a signer-independent split and PUCP-DGI30 a stratified random split (as described in the WholeBodyPose README). - DGI_PUCP38: all 931 videos of PUCP-DGI305 with their 38 classes (DGI_PUCP30 keeps 30 of them) and
the 651/140/140 split used in SignERT experiments. It was extracted on 2026-11-20 with
wholebodypose-build-islr --dataset pucp_dgi --subset 38 --preset icmv2025, in the sample order of the original 96-keypoint files. Against those files, itsSIGNERT_96columns agree with a median difference of 8.5e-9 (95th percentile 2.8e-8, max 0.11) in normalized coordinates, with identical labels, video names and frame counts for all 931 clips. It is a separate extraction, so it is not bit-identical to DGI_PUCP30 on the shared videos.
Loading
import json
import h5py
import numpy as np
from huggingface_hub import hf_hub_download
repo = "CristianLazoQuispe/pose-action-recognition"
h5_path = hf_hub_download(repo, "ISLR/WLASL/WLASL100/WLASL100_135-Train.hdf5", repo_type="dataset")
map_path = hf_hub_download(repo, "ISLR/WLASL/WLASL100/wlasl_100_maplabels.json", repo_type="dataset")
gloss_to_id = json.load(open(map_path))["id_to_label"] # gloss -> id (see "Label maps")
samples = []
with h5py.File(h5_path, "r") as f:
for key in f: # "0", "1", ...
g = f[key]
xy = g["data"][()].transpose(0, 2, 1) # (T, 135, 2), normalized x, y
gloss = g["label"][()].decode()
size = (int(g["width"][()]), int(g["height"][()]))
samples.append((xy.astype(np.float32), gloss_to_id[gloss], size))
xy, y, (w, h) = samples[0]
pixels = xy * np.array([w, h], dtype=np.float32) # pixel coordinates (undetected points stay at 0, 0)
coco133 = xy[:, :133] # drop neck and pelvis -> plain COCO-WholeBody
To download a whole subset:
from huggingface_hub import snapshot_download
snapshot_download(
"CristianLazoQuispe/pose-action-recognition",
repo_type="dataset",
allow_patterns=["ISLR/LSA64/*"],
local_dir="pose-action-recognition",
)
License
- Attribution: cite WholeBodyPose and the source dataset of every subset you use (Citation).
- NonCommercial: no commercial use.
- ShareAlike: share derived data under CC BY-NC-SA 4.0.
- The terms of each source dataset (below) also apply to its subset.
Please do not attempt to re-identify the signers.
Source datasets and their terms
These files are derived data: keypoints computed from the videos of the datasets below. The videos are not redistributed. The terms of each source dataset still apply to its subset; read them before use, and cite the original dataset.
| Subset | Source dataset | Terms of the source dataset |
|---|---|---|
| WLASL* | WLASL (Li et al., WACV 2020) | Computational Use of Data Agreement (C-UDA); "intended for academic and computational use only. No commercial usage is allowed." |
| MSASL* | MS-ASL (Vaezi Joze & Koller, BMVC 2019) | See the original dataset terms. |
| ASLCitizen* | ASL Citizen (Desai et al., NeurIPS 2023 Datasets and Benchmarks) | See the original dataset license; for commercial use the authors ask to contact ASL_Citizen@microsoft.com. |
| LSA64 | LSA64 (Ronchetti et al., CACIC 2016) | CC BY-NC-SA 4.0: non-commercial, attribution required, derivatives under the same license. |
| DGI_PUCP30, DGI_PUCP38 | PUCP-DGI: Peruvian Sign Language recordings by the Linguistics Department of the Pontificia Universidad CatΓ³lica del PerΓΊ (RodrΓguez MondoΓ±edo and Arnaiz, 2015), as used in PeruSIL (Bejarano et al., sign-lang@LREC 2022) | See the original dataset terms. |
Citation
You must cite both WholeBodyPose (the pose extraction and these files) and the source dataset of every subset you use. Citing only one of them does not meet the license.
(1) WholeBodyPose (always):
@inproceedings{lazo2026wholebodypose,
title = {WholeBodyPose: a unified end-to-end framework for sign language recognition and pose-based training data},
author = {Lazo-Quispe, Cristian and Castro-Cruz, Renato and Bejarano, Gissella and Salazar-Espinosa, Mauricio},
booktitle = {Eighteenth International Conference on Machine Vision (ICMV 2025)},
volume = {14114},
pages = {69--79},
year = {2026},
organization = {SPIE},
doi = {10.1117/12.3096312}
}
(2) The source dataset of each subset you use:
| Subset | Cite |
|---|---|
| WLASL100 / 300 / 2000 | li2020word |
| MSASL100 / 200 / 500 / 1000 | joze2019msasl |
| ASLCitizen100 / 300 / 1000 / 2731 | desai2023aslcitizen |
| LSA64 | ronchetti2016lsa64 |
| DGI_PUCP30 | rodriguez2015lsp and bejarano2022perusil |
@inproceedings{li2020word,
title = {Word-level Deep Sign Language Recognition from Video: A New Large-scale Dataset and Methods Comparison},
author = {Li, Dongxu and Rodriguez, Cristian and Yu, Xin and Li, Hongdong},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
pages = {1459--1469},
year = {2020}
}
@inproceedings{joze2019msasl,
title = {MS-ASL: A Large-Scale Data Set and Benchmark for Understanding American Sign Language},
author = {Vaezi Joze, Hamid Reza and Koller, Oscar},
booktitle = {British Machine Vision Conference (BMVC)},
year = {2019}
}
@inproceedings{desai2023aslcitizen,
title = {ASL Citizen: A Community-Sourced Dataset for Advancing Isolated Sign Language Recognition},
author = {Desai, Aashaka and Berger, Lauren and Minakov, Fyodor and Milano, Nessa and Singh, Chinmay and Pumphrey, Kriston and Ladner, Richard and Daum{\'e} III, Hal and Lu, Alex X. and Caselli, Naomi and Bragg, Danielle},
booktitle = {Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track},
pages = {76893--76907},
year = {2023},
doi = {10.52202/075280-3360}
}
@inproceedings{ronchetti2016lsa64,
title = {LSA64: A Dataset of Argentinian Sign Language},
author = {Ronchetti, Franco and Quiroga, Facundo and Estrebou, C{\'e}sar and Lanzarini, Laura and Rosete, Alejandro},
booktitle = {XXII Congreso Argentino de Ciencias de la Computaci{\'o}n (CACIC)},
year = {2016}
}
@misc{rodriguez2015lsp,
title = {Repositorio Lengua de Se{\~n}as Peruana},
author = {Rodr{\'i}guez Mondo{\~n}edo, Miguel and Arnaiz, Alexandra},
howpublished = {Pontificia Universidad Cat{\'o}lica del Per{\'u}},
year = {2015}
}
@inproceedings{bejarano2022perusil,
title = {PeruSIL: A Framework to Build a Continuous Peruvian Sign Language Interpretation Dataset},
author = {Bejarano, Gissella and Huamani-Malca, Joe and Cerna-Herrera, Francisco and Alva-Manchego, Fernando and Rivas, Pablo},
booktitle = {Proceedings of the 10th Workshop on the Representation and Processing of Sign Languages (sign-lang@LREC 2022)},
pages = {1--8},
year = {2022},
publisher = {European Language Resources Association (ELRA)}
}
For the 133-keypoint layout, also cite COCO-WholeBody:
@inproceedings{jin2020whole,
title = {Whole-Body Human Pose Estimation in the Wild},
author = {Jin, Sheng and Xu, Lumin and Xu, Jin and Wang, Can and Liu, Wentao and Qian, Chen and Ouyang, Wanli and Luo, Ping},
booktitle = {ECCV},
year = {2020}
}
- Downloads last month
- 170