armin@csri.toronto.edu (Armin Haken) (03/23/88)
AI Seminar: 11AM, Thursday March 31, Sandford Fleming 1101 (Note that the speaker is a candidate for a faculty position.) BAYESIAN MODELING OF UNCERTAINTY IN EARLY VISION Richard Szeliski Computer Science Deparment Carnegie Mellon University Many of the problems in early vision, such as stereo, motion, shape from shading and surface interpolation, have recently been formalized using variational calculus, regularization and Markov Random Fields. These techniques result in well defined computational problems and lead to algorithms that can incorporate sparse or noisy data from a variety of sensors. The focus to date has been on finding algorithms that yield single optimal estimates, without considering the resulting uncertainty in these estimates. The work in my thesis starts by interpreting the smoothness constraints used in regularization as defining a probabilistic prior model. This allows us to compute the uncertainty in the posterior estimate, and to model probability distributions over dense fields (such as depth maps) in a succinct fashion. The uncertainty can be used directly in applications such as robot navigation. The estimate thus obtained can also be combined with new measurements to yield an improved estimate whose uncertainty is reduced over time. This approach, which is an extension of the Kalman filter to two-dimensional depth maps, has been used to design an incremental (on-line) depth-from-motion algorithm. Our results show that this incremental algorithm performs as well as wide-baseline stereo, and is simpler to implement.