pedroodb/glosl-lsat

GloSL · LSA-T — Continuous Argentinian Sign Language

It is part of the GloSL effort to unify continuous Sign-Language-Translation corpora into one common format (normalized pose keypoints + aligned text/gloss).

• Sign language: Lengua de Señas Argentina (LSA) (aed, ISO 639-3)
• Spoken language: Spanish (es)
• Task: continuous sign-language translation
• License: repo MIT (code only); dataset data unlicensed
• Pose redistribution (reupload_ok): public
• Cropped-video redistribution (video_rehost): permitted — see note below
• Video provenance / permission: the source video originates from the Canal Sordos Argentina (CN Sordos) YouTube channel and is republished here WITH the channel's explicit permission. Pose + video are hosted publicly on this basis; no further author authorization is required.

Citation / source

• Paper: Dal Bianco et al., LSA-T, IBERAMIA 2022 (arXiv:2211.15481)
• Source page: https://midusi.github.io/LSA-T/

Contents

• annotations.csv — GloSL common-core: id, split, text (+ gloss/signer/start/end where present)
• poses/<id>.pose — MediaPipe-Holistic keypoints (543 pts: 33 body + 468 face + 21+21 hands) in the ecosystem-standard .pose container. Normalization (glosl-norm-v1) is applied at load time, not stored.
• card.json — machine-readable GloSL dataset card.

Splits

{ "dev": 1354, "train": 5413, "test": 1692 }

Pose provenance

Reused from the group's harmonized slt_datasets MediaPipe-Holistic .npy (see GloSL KB/12); repackaged to .pose by glosl-datasets. Estimator: mediapipe-holistic, dims=3.

Video

LSA-T videos are the dataset's own per-utterance signer clips (CN Sordos, signer-centric; ~1 person/frame at 1920x1080). Stored at source framing; not additionally YOLO-bbox-cropped in v1. Republished with the explicit permission of the CN Sordos (Canal Sordos Argentina) YouTube channel.