Target-focused video stabilization for human computer interaction

As an important task of video enhancement, a lot of video stabilization methods have been proposed. However in many real-world applications, especially the systems with human-computer interactions, existing methods only remove camera motion in stabilized frames, the remaining object motion will also lead to deviations in manual inputs. In this paper, we collect practical hand drawn bounding boxes which have been shown to contain serious errors. Then we propose a target-focused video stabilization method consisting of a proposal-based detection component and a tracking-based motion estimation component. The experiments demonstrate our method can remove camera jitter and target motion simultaneously, and also offer users a friendly and effective way to draw accurate target regions.