LSA-T: A Continuous Argentinian Sign Language Dataset

LSA-T is the first continuous dataset for Argentinian Sign Language (LSA). Its videos come from the CN Sordos YouTube channel, which publishes videos in LSA with Spanish subtitles. Each clip covers one subtitle, so the dataset pairs signed sentences with their Spanish translation.

For each clip the dataset includes the keypoints of every person in the frame and, when there is more than one, the person inferred to be signing.

Versions

The version described in the paper has 14,880 clips by 103 signers, with a vocabulary of 14,239 Spanish words. The current release has 8,459 clips, with metadata on the number of people in each clip and the inferred signer.

Files

• meta.csv: one row per clip, with its label and signer metadata.
• clips.7z: the video clips, named by their id in meta.csv.
• keypoints.h5: an HDF5 file with one group per clip and one subgroup per person (signer_i).

Download links and a PyTorch loader are on the LSA-T project page and in the GitHub repository. The clips can also be browsed and validated on Seni.ar.

Samples

Clips from the dataset, with the detected keypoints drawn over each person and a box around the inferred signer.

@inproceedings{dalbianco2022lsat,
  title={{LSA-T}: The First Continuous Argentinian Sign Language Dataset for Sign Language Translation},
  author={Dal Bianco, P. A. and Ríos, G. G. and Ronchetti, F. and Quiroga, F. M. and Stanchi, O. A. and Hasperué, W. and Rosete, Alejandro},
  booktitle={Advances in Artificial Intelligence -- IBERAMIA 2022},
  pages={293--304},
  year={2022},
  organization={Springer},
  doi={10.1007/978-3-031-22419-5_25}
}