Hi Gang,
This week we will return to our regularly schedule tea talks with a talk by David Rim. Title and abstract are below. Hope to see you there.
Time: 14h00, Thursday August 18th Location: LISA lab (AA3256)
Title: Learning Identity-Invariant Features for Expression Recognition with 3D CGI Models for Semi-Supervised Learning
Abstract:
I plan to propose a method for learning features for expression recognition by leveraging work done in the graphics community for modeling and generating realistic human faces and related motion-capture technology. I will demonstrate that these models provide expression representations which are far more compact and more importantly, easier to interpret and code than Facial Action Coding System Action Units. I will demonstrate a labeling process which I believe will lead to more accurate labels, and show initial experiments using simple linear models for feature learning.
I will propose a convolutional network model to learn better features (filter maps) across models for different "identities" as well as using unlabeled human faces. Finally, I will propose a semi-supervised model which can use these features to learn an inductive function which maps 2D facial images to the expression representation using these learned features inductively. I hope a useful discussion will follow.
Cheers, Aaron