Supervisors: Dr. Zhiwu Huang, Dr. Danda Pani Paudel, Prof. Luc Van Gool
Automatic analysis of facial expressions is an important problem in many fields such as sociable robotics, driver fatigue surveillance, medical treatment, and other human-computer interaction systems. In this work, we aim to answer the question of “which representation suits best for facial expression generation as well as recognition?” In particular, we study the utility of three independent affectional dimensions, namely pleasure, arousal and dominance, to describe the perception of human emotion in a continuous manner. To do so, we adopt a hybrid approach of conditional and cycle GAN for image-to-image transfer. To justify our discovery, two different input types from action units and affectional dimensions are tested. The first type of action units produce robust results, preserving identities with no significant artifacts. However, we argue that the meaningful manipulation of these high dimensional action units is cumbersome. Therefore, we suggest to make use of only three affectionate dimensions for facial animation, showing that they are not only simple, intuitive and sufficient, but also offer better qualitatively and quantitatively results.