Advancements in sign language processing technology hinge on the availability of extensive, reliable datasets, comprehensive instructions, and adherence to ethical guidelines. To facilitate progress in gesture recognition and translation systems and to support the Azerbaijani sign language community we present the Azerbaijani Sign Language Dataset (AzSLD). This comprehensive dataset was collected from a diverse group of sign language users, encompassing a range of linguistic parameters. Developed within the framework of a vision-based Azerbaijani Sign Language translation project, AzSLD includes recordings of the fingerspelling alphabet, individual words, and sentences. The data acquisition process involved recording signers across various age groups, genders, and proficiency levels to ensure broad representation. Sign language sentences were captured using two cameras from different angles, providing comprehensive visual coverage of each gesture. This approach enables robust training and evaluation of gesture recognition algorithms. The dataset comprises 30,000 meticulously annotated videos, each labeled with precise gesture identifiers and corresponding linguistic translations. To facilitate efficient usage of the dataset, we provide technical instructions and source code for a data loader. Researchers and developers working on sign language recognition, translation, and synthesis systems will find AzSLD invaluable, as it offers a rich repository of labeled data for training and evaluation purposes.
Read full abstract