SignBD-Word: Video-Based Bangla Word-Level Sign Language and Pose Translation

Bangla sign language (BdSL) is a complete and independent natural sign language with its own linguistic characteristics. While there exists video datasets for well-known sign languages, there is currently no available dataset for word-level BdSL. In this study, we present a video-based word-level dataset for Bangla sign language, called SignBD-Word, consisting of 6000 sign videos representing 200 unique words. The dataset includes full and upper-body views of the signers, along with 2D body pose information. To evaluate the dataset, a number of deep learning-based algorithms for video classification and pose translation are used. The experimental results reveal that the recently developed SlowFast and I3D methods achieve higher accuracy for RGB and body pose data respectively, compared to the traditional attention-based approach. Furthermore, the experiments indicate that pix2pixHD outperforms CycleGAN and SPADE for human- to-human sign pose translation from bodypose data.