Machine perception primitives for real time social interaction

Fortenberry, B., Bartlett, M.S., Roddey, J.C., Fasel, I., Marks, T., Littlewort, G., Chenu, J., Movellan, J.R.

18th Conference on Neural Information Processing Systems
Vancouver, Canada. December 13-16, 2004.


We present four demonstrations of primitives for real-time social interaction. (1) Facial Expression Recognition: using a texture-based recognition system that employs Gabor wavelet decomposition and support vector machines, we can recogniz e 7 expressions of basic emotion as well as 18 action units from the Facial Acti on Coding System (FACS). (2) Real-Time Unsupervised Object Learning: we present a generative approach for performing unsupervised learning and tracking of faces . (3) G-flow: using a stochastic filtering algorithm based on a conditionally Ga ussian generative model, we track in 3D the rigid and non-rigid motion of a face from 2D video. (4) Infomax Controller for Social Contingency Detection: we pres ent a humanoid robot that implements an optimal Bayesian controller. The control ler schedules the robot's behavior in real time to maximize the speed and accura cy of detection of the presence of people.