Scaling up ConvAtt for Sign Language Recognition

Abstract

Sign language is crucial for communication within the deaf community, making Sign Language Recognition (SLR) essential for bridging the gap between signers and non-signers. However, SLR models often face challenges due to limited data availability and quality. This paper investigates various data augmentation and regularization techniques to enhance the performance of a lightweight SLR model. We focus on recognizing signs from the French Belgian Sign Language using a novel model architecture that integrates convolutional, channel attention, and selfattention layers. Our experiments demonstrate the effectiveness of these techniques, achieving a top-1 accuracy of 49.99% and a top-10 accuracy of 83.19% across 600 distinct signs.