[May 3rd 2018] LabMeeting: Modeling stylized character expression via deep learning

Lisa Graziani (DIISM, University of Siena)

May 3, 2018 – 9:30 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI

DeepExpr is a perceptual model that allows an accurate retrieval of stylized character expressions given a human expression query. It can be helpful in animation to create desired character expressions driven by human facial expressions.

First two Convolutional Neural Networks were trained independently to recognize the expression of humans and stylized characters. Then a transfer learning technique was used to learn the mapping from humans to characters to create a shared embedding feature space.

A database of labeled facial expressions (6 basic emotions + neutral) for six stylized characters was created.

 |  Category: Seminars