V-Librasil - A New Dataset with Signs in Brazilian Sign Language (Libras)

This work presents a new labeled dataset of videos with native and professional interpreters articulating words and expressions in Libras (Brazilian Sign Language). We used a methodology based on related studies, the support of the team of articulators, and the existing datasets in the literature. We also created a correlation of signals with the principal signal dataset in the literature, the American Sign Language Lexicon Video Dataset - ASLLVD, to enable an association between the signals in Libras and ASL, and promote the development of recognition and translation systems between the two sign languages. The V-Librasil dataset consists of a Portuguese-Libras compilation. The Portuguese partition is formed by 1364 terms extracted and translated from the ASLLVD dataset. These terms were communicated in the Brazilian Sign Language by three articulators, totaling 4089 signs. Furthermore, the signals were filmed in a chroma key environment, allowing the simulation of different backgrounds after recording the signals. The dataset is publicly available on a structured platform https://libras.cin.ufpe.br/. This dataset will allow training and validation of different models based on machine learning that can recognize and translate signals from one language to another since the literature lacks datasets in Libras this scope, quality, and quantity