Machine learning approach to automate facial expressions from physical activity

We propose a novel approach based on machine learning to simulate facial expressions related to physical activity. Because of the various factors they involve, such as psychological and biomechanical, facial expressions are complex to model. While facial performance capture provides the best results, it is costly and difficult to use for real-time interaction during intense physical activity. A number of methods exist to automate facial animation related to speech or emotion, but there are no methods to automate facial expressions related to physical activity. This leads to unrealistic 3D characters, especially when performing intense physical activity. This research highlights the link between physical activity and facial expression, and to propose a data-driven approach providing realistic facial expressions, while leaving creative control. First, biological, mechanical, and facial expression data are captured. This information is then used to train regression trees and support vector machine (SVM) models, which predict facial expressions of virtual characters from their 3D motion. The proposed approach can be used with real-time, pre-recorded or key-framed animations, making it suitable for video games and movies as well