Libras-10 Dynamic Signs Database

P1 - letter H P2 - day P3 - again P4 - yours P5 - letter J
P6 - night P7 - enter P8 - all P9 - start P10 - course

The Libras-10 Dynamic Signs Database is composed by samples of a set of 10 classes representing dynamic signs of Libras, which were selected by Libras (Brazilian Sign Language, BSL) specialists. The project was carried out by researchers and students from Federal University of Alagoas, Brazil. The dynamic signs were chosen to satisfy the following requirements: 5 one-handed signs and 5 two-handed signs; all signs should be very distinct, to increase diversity, except for two pairs of similar signs which are harder to discriminate ("night" and "enter"), to better evaluate the robustness of the models. For each class, we collected approximately 300 examples of depth videos. The number of frames per example varies, and the longest example contains 40 frames. The database was originally collected to evaluate the methodology presented in the paper "Lucas Amaral, Givanildo L. N. JĂșnior, Tiago Vieira and Thales Vieira, 'Evaluating deep models for dynamic brazilian sign language recognition'. In Iberoamerican Congress on Pattern Recognition 2018. Springer, Cham".

The database can be downloaded here.

Structure of the database folders: Classes/Samples/im#.png , where:

  • the root contains one folder for each trained sign class;
  • the Classes folders contain subfolders with trained samples of its corresponding trained sign;
  • the Samples subfolders contain the depth images from each video, in png format.

    Our database is free to use for research purposes. Please cite our database in your work:

    title={Evaluating deep models for dynamic brazilian sign language recognition},
    author={Lucas Amaral, Givanildo L. N. J{\'u}nior, Tiago Vieira and Thales Vieira},
    booktitle={Iberoamerican Congress on Pattern Recognition},

    For questions, please contact Prof. Thales Vieira.