Real-time content aware resizing of video
MetadataShow full item record
In this thesis, we propose a new method for content-aware resizing of videos in real- time. Our approach consists of two steps. First, we compute a set of non-salient pixels in linear time which, when being removed or duplicated, do not alter the general appearance of the video. This is an extension of Avidan and Shamir's  greedy seam-carving approach to video. Second, we generate a new representation of the video, so called multi-view videos that allow us to resize the video in real-time, i.e. while being watched. This representation can be computed very effciently, the complexity is linear in the number of frames and linear in the number of pixels in a video. Our technique works on a broad variety of videos and is computationally inexpensive enough to be executed by a vast range of devices. We compare our technique to our own implementation of a current state-of-the-art approach and show several convincing results obtained by our technique.