Hybrid optimization for automatic large-scale video annotation

This project aims to develop modern optimization techniques in the area of computer vision for theapplication of automatic video annotation. A demonstrator for this research is the fully automatic annotation of a large volume of roadside (street-view) video. We propose a novel hybrid optimization scheme whichcombines the advantages of robust continuous optimization (particularly L1-norm based) methods, with discrete optimization (e.g. solution of Markov Random Fields). Emphasis of the project is given not only to rigorous theoretical analysis of the optimization techniques, but also to significant promises it holds in solving many practical computer vision problems.