Sign Language provides the means of conveying messages for deaf and mute people. Effective communication with the masses is a great challenge for the deaf and mute community, as Sign Language is not commonly understood. Many researchers have done numerous works in foreign language datasets like English, French, Japanese, etc. However, for Bangla, one of the most widely spoken languages, much significant work has not been done yet. Most of the works on Bangla Sign Language are executed on small datasets and report satisfactory performance. However, when small datasets are evaluated from the perspective of generalizability, particularly when using deep learning based solutions, these models fail to reproduce to the same performance. Therefore, this paper poses inter-dataset evaluation as the main evaluation criteria and evaluates several deep learning based models. This evaluation is done for Bangla by leveraging two popular datasets of Bangla Sign Language. Unsurprisingly, the inter-dataset performance is inferior, and several approaches to improve are explored and documented, including the use of angular margin based loss functions. The results demonstrate the importance of such an evaluation and also show that one of the proposed approaches shows promising performance, albeit with significant room for improvement. This raises the need for a standardized dataset to overcome this issue of generalization for real-life applications, as well as the need to encourage future works to concentrate on challenging evaluations instead of pursuing deceptively good intra-dataset performance.

No abstract