Final Report Summary - VISLIM (Visual Learning and Inference in Joint Scene Models)
We have developed a variety of tools for inference and learning in complex scene models, such as discriminative approaches that can cope with important structural parameters that are only available at test time, or are estimated jointly at test time. Moreover, we showed that these discriminative approaches can be directly derived from a generative model combined with certain inference algorithms. We have investigated inference tools that can cope with the challenging case of mixed discrete/continuous joint scene representations, probabilistic methods for jointly estimating the uncertainty including highly practical and easy-to-use approaches, and also proposed novel tools and methods for obtaining high-quality data needed to train discriminative approaches. Our research on joint scene models focused on complex interactions between segmentations, motion, occlusions, and geometry, or between appearance, motion, and segmentations. Applications addressed by our work include motion blur estimation and localization, image deblurring, 3D scene flow estimation, optical flow estimation, joint shape and lighting estimation, as well as monocular scene analysis. In several areas leading results have been achieved on agreed-upon benchmarks setting a new state of the art.