Multi-View Bangla Sign Language Recognition: A New Word-Level Video Dataset

Sign language recognition is essential for overcoming communication barriers, especially for those with verbal challenges. However, recognizing sign language presents various challenges, including shared gestures, lighting, bodily poses, and environmental differences. The scarcity of a comprehensive Bangla sign language video dataset exacerbates these challenges, especially for deep learning-based algorithms. To address this gap, we develop the MVBSL-W50, a multi-view Bangla sign language dataset encompassing 50 lexically isolated signs across 13 categories. We also design a model based on human pose information, achieving an 89.69% accuracy. We conduct experiments to evaluate the model's performance against angular variations and lighting conditions, emphasizing its robustness and applicability in real-world settings. Our model is further evaluated on the Indian Lexicon Sign Language Dataset (INCLUDE), where it achieve an accuracy of 96.60%. This significant improvement underscores the effectiveness of our approach in Bangla Sign Language (BSL).