Séminaire LATECE: «Sign Language Recognition and Translation»
Bonjour à tous,
Vous êtes cordialement invités à assister en ligne au premier séminaire du LATECE de la session d'automne 2020.
Heure et date : Le 16 septembre 2020, 12:30
Lieu : réunion Zoom : https://uqam.zoom.us/j/92718409435
Conférencier : Fares Ben Slimane, Étudiant à la maîtrise sous la direction du professeur Mohamed Bouguessa
Titre : Sign Language Recognition and Translation
Résumé : Besides hand gestures, sign language simultaneously uses different components to convey a certain message. For example, the orientation of the fingers, the movements of the arms or body as well as the facial expressions. Sometimes a specific component can play a major role in changing the meaning of the sign or it may not be required to interpret a sign at other times. Accordingly, it is essential for a recognition system to use only the relevant information for sign prediction. In this context, we propose the Sign Transformer Network, an attentional network for both tasks: Continuous Sign Language Recognition and Sign Language Translation. It takes as input a sequence of images which designates the sign to translate and produces a coherent textual translation in a spoken language. Our system is based on the new neural architecture, Transformer Network, which has the capacity to efficiently discover and learn the spatio-temporal information of continuous data. We show that by merely using self-attention for temporal modelling, we outperform nearly all previous studies, proving the superiority of self-attention over traditional recurrent based networks.