Disney Research has developed a facial-capture system that uses a sample of actors' recordings to synthetically generate the data needed to train the system.
Unlike traditional machine-learning techniques that rely on capturing actors displaying a variety of facial expressions under numerous lighting and camera conditions to build a dataset, the new method would enable datasets 10 to hundreds of times smaller to be used without affecting facial capture accuracy.
Researchers use a multi-camera setup to record about 70 facial expressions under uniform lighting conditions, generating a movable model of the actor's face. The model is then used to produce the synthetic training data tailored for various conditions similar to those expected on the actual set.
The researchers found focusing on expression and changes in illumination, instead of variations in camera perspectives, achieved the best accuracy. They say the technique reduces training image counts by one to two orders of magnitude with no visible loss of accuracy.
"By reducing the amount of facial imagery necessary to train these systems, our team has taken a big step toward increasing the flexibility and efficiency of this approach," says Disney Research's Markus Gross.
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found