Bangla Sign Language Generation From Videos

Abstract

Sign language serves as a crucial communication tool for the Deaf community, representing a fully developed linguistic system. As research into sign language

via deep learning expands, the need for an efficient and robust toolset for sign lan- guage data collection becomes increasingly important. This abstract introduces

the Sign Language Data Collection Toolset (SLDCT), which is designed to stream- line the processes of collecting and annotating sign language data. The SLDCT

includes various interconnected components such as user registration, video record- ing, image extraction, keypoint extraction, and the selection of multiple regions of

interest. The video recording software, a central feature of the toolset, facilitates the easy recording of sign language performances. A significant challenge in sign language data collection is the annotation process, which is often time-consuming

and requires detailed analysis by experts. To mitigate this issue, the SLDCT in- tegrates annotation tools that allow for real-time annotation during the recording

process, thereby merging the recording and annotation phases. This integration helps researchers annotate data more efficiently. One notable feature of the toolset

is the ability to collect keypoints. The toolset can extract frames from videos, cap- ture facial landmarks, hand poses, and other keypoints, and save this information

in a .csv file. These keypoints are valuable for training lightweight deep-learning models for various research applications. Additionally, the toolset allows for the capture of multiple regions of interest from video frames, with the bounding box coordinates saved as metadata alongside the dataset.