OkkhorNama: BdSL Image Dataset For Real Time Object Detection Algorithms

Abstract

In recent years, lots of researches are being conducted to interpret Bangladeshi Sign Language (BdSL) to the means that general people can communicate with people having a hearing impairment and reduce the verbal gap between them. Computer Vision is playing a vital role in this regard by developing a sustainable system to understand the signs for machine translations. To obtain optimal performance, along with the state-of-the-art CNN model, the requirement of a high-quality sign language dataset cannot be foreseen. In this paper, we have introduced a new image dataset OkkhorNama for Fingerspelled Bangladeshi Sign Language including all 46 signs with images over 12K. In each of the images, bounding boxes are carefully annotated and labeled. OkkhorNama contains images of high resolution, good quality, and adequate variation making it ideal to train object detection and localization algorithms that would perform well on real-world applications. The OkkhorNama dataset is compared with other datasets where OkkhorNama significantly outperforms other datasets in number and trained model performance. The dataset is publicly available for future research and development.