DiPASL the Different Perspective ASL

Sign Language Recognition (SLR) has gained significant popularity in the field of computer vision. However, existing methods for recognizing hand sign language predominantly assume visuals captured from the viewpoint of the viewer while this paper understands American Sign Language (ASL) from the user-centric perspective which accurately represents how speciallyabled individuals communicate with others. It is fundamental to consider this least studied approach that will prove to be most impactful in the development of highly user-friendly real-time pipelines. We developed two datasets to conduct our study, DiPASL-T900 with nonprocessed RGB images and DiPASL-S900 with skeleton mappings. Extensive comparisons of performances are done for letter-by-letter finger-spelling using five Convolutional Neural Network (CNN) models with structured hyperparameter tuning. Additionally, certain hand signs are designated as stop symbols, enabling the grouping of classified letters into meaningful words, which are then translated into speech. Different technological hooks like cropping (to reduce the focus of the model to a very small region of interest), augmentation, and mappings are used to achieve creditable accuracy in real-time. First-of-its-kind qualitative and traditional quantitative metrics are utilized to present proof of concept in the feasibility of the approach. Thereby, this work presents a novel perspective on SLR with a more inclusive and user-friendly system for sign language users. The potential for real-time improvements holds promise for enhancing the usability and practicality of SLR in various contexts.