CSE Fac Candidate, Feng Liu

In a utopia of visual media, everyone can easily create, enjoy and share high-quality visual media. My research brings us closer to this dream by developing novel tools for authoring visual media, technologies for effectively presenting visual media on heterogeneous devices, and methods for leveraging large-scale user-contributed media collections for visual media synthesis and understanding.

In this talk, I will present my work on developing visual media authoring tools. My focus will be the development of a video stabilization approach that can transform an amateur video so that it appears as if it were created by a professional filmmaker along a directed camera path. First, I will describe how we use an understanding of human perception to motivate our design of the first practical video stabilization system capable of these dramatic transformations via a 3D approach. Then I will describe how we remove 3D reconstruction from our method by using a subspace constraint on video feature trajectories. This leads to a solution that is robust, efficient and scalable enough to be deployed in commercial products. Finally, I will briefly describe my work on large-scale web media collections, demonstrating how these collections can be used to provide sources for panorama synthesis.