Motion (Occlusion) Boundaries

Research Overview

One of the most significant remaining problems in early motion estimation concerns occlusion boundaries, where the two key assumptions of current optical flow methods, brightness conservation and motion smoothness, are typically violated. Michael Black and I recently formulated a probabilistic solution to this problem with a hybrid state-space model and a particle filter for approximate inference. The state space model included a discrete random variable to represent different motion classes (e.g., smooth motion, or discontinuous motion), and continuous variables to represent the parameters of each motion class. This work won Honorable Mention for the Marr Prize (runner-up for best paper) at the International Conference on Computer Vision in 1999.

Although these initial experiments produced encouraging results the method proved to be unreliable. Oscar Nestares and I have since improved the method by introducing a random field of local neighborhoods to encourage spatiotemporal continuity of the inferred surface boundaries, and an empirical edge-based likelihood function to improve boundary localization.

Related Publications

Return to David Fleet's home page.