Advanced Models for Brightness Changes and Confidence Measures in Motion Estimation

Claudia and Daniel Kondermann
(University Heidelberg, Germany )

In this talk we describe a framework for application-adaptive models in optical flow estimation. Starting out from some typical problems, we first sketch several approaches to solve for the optical flow and additional parameters, explicitly addressing physically motivated brightness variation models. These can be incorporated into local or global motion estimation techniques. Motion models cannot only be used for the estimation of the optical flow. In conjunction with confidence measures, they can also be used for selecting highly accurate flow vectors. We will present novel approaches to formulating such confidence measures. Here, the underlying motion models are learned from flow fields typically arising in specific applications. Based on confidence measures corrupted flow fields can be thinned out and, if necessary, reconstructed to obtain dense motion fields. Finally, we discuss some typical application scenarios were these methods are useful.