Research 3
Real-time vision for human-machine interaction
A powerful approach to computer vision problems has come from deconstructing traditional computer vision algorithms and re-stating them in terms of probabilistic, generative frameworks. This allows us to make explicit the basic assumptions (which may have been hidden before), allows derivation of optimal inference algorithms, and facilitates integration with other sources of information that may not have been considered initially. This has resulted in a variety of novel applications such as real-time detection of facial features (e.g., eyes) and facial action-units (e.g., eye-blinks), and is led to new algorithms such as Segmental Bolzmann Fields and the BEV project.
These systems have allowed us to develop social robots that can interact with people in a more natural way than via keyboard, mouse and screen.