Pose-Based Deep Learning Approaches for Recognizing Isolated Signs in Greek Sign Language

Summary

This thesis explores the task of isolated sign recognition in Greek Sign Language (GSL) using deep learning. GSL, like many sign languages, lacks large-scale annotated datasets, making auto-matic recognition a challenging problem. To address this, we use the publicly available GSL RGB+D dataset, which contains annotated video recordings captured with an Intel RealSense depth camera. We implement and evaluate three distinct neural architectures: a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a Graph Convolu-tional Network (GCN). Each model is designed to handle different characteristics of sign lan-guage data, visual, temporal, and spatial. Our experiments, conducted on the isolated gloss subset of the dataset, show that the LSTM model achieves the highest overall accuracy, while the CNN and GCN models demonstrate strength in specific categories. These findings underline the im-portance of temporal and structural information in sign recognition. This work contributes a com-parative study of recognition models tailored to Greek Sign Language and highlights their poten-tial in low-resource language contexts.