Non-acted multi-view audio-visual dyadic interactions. Project master thesis: multitask learning for facial attributes analysis

Abstract

[en] In this thesis we explore the use of Multitask Learning for improving performance in facial attributes tasks such as gender, age and ethnicity prediction. These tasks, along with emotion recognition will be part of a new dyadic interaction dataset which was recorded during the development of this thesis. This work includes the implementation of two state of the art multitask deep learning models and the discussion of the results obtained from these methods in a preliminary dataset, as well as a first evaluation in a sample of the dyadic interaction dataset. This will serve as a baseline for a future implementation of Multitask Learning methods in the fully annotated dyadic interaction dataset.