Evaluating deep models for dynamic brazilian sign language recognition

Abstract

We propose and investigate the use of deep models for dynamic gesture recognition, focusing on the recognition of dynamic signs of the Brazilian Sign Language (Libras) from depth data. We evaluate variants and combinations of convolutional and recurrent neural networks, including LRCNs and 3D CNNs models. Experiments were performed with a novel depth dataset composed of dynamic signs representing letters of the alphabet and common words in Libras. An evaluation of the proposed models reveals that the best performing deep model achieves over 99% accuracy, and greatly outperforms a baseline method.