VDzSL: A synthetic video dataset for Algerian sign language using 3D avatars

Video datasets are crucial for advancing communication technologies for deaf and hard-of-hearing individuals. Despite that, extensive datasets are not available for the majority of sign languages due to the ample work required to capture, clean, organize, and publish them. This paper introduces the Video Dataset for Algerian Sign Language (VDzSL), the largest video dataset for Algerian Sign Language. To ensure demographic diversity, VDzSL utilizes four different avatars to animate the signs and records them from five distinct camera angles, employing polar coordinates to ensure consistency while capturing varying horizontal and vertical perspectives. With 415 signs, our dataset has a 99.5% coverage of the signs included in 3DZSignDB's SiGML dataset, and 26.6% coverage of the official ALGSL dictionary provided by the Algerian Ministry of Solidarity. Our dataset contains 8300 video files totaling 3 h, 11 min, and 43 s of synthetic videos provided at a 498×498 pixel resolution and an average frame rate of 27 frames per second across the entire dataset. The dataset is primarily aimed at training, testing, and benchmarking machine learning models, facilitating transfer learning and comparative analyses, as well as developing learning tools and accessibility applications.