Sharing video annotations

This paper describes an approach for locating annotations generated in one video and properly placing them in a second, modified version of the same video. We focus on modifications that standard Television (TV) broadcasts may experience, including content insertion and deletion (i.e., commercials), format conversions and different start/end recording times. To overcome these modifications, we propose content-based video timelines. We identify the position of a frame in a long video stream with an accuracy of one frame based on its content, without using embedded time codes. To make this approach feasible, we use a compact representation of a video frame, which we call a "fingerprint". Fingerprints capture small temporal variations within shots and therefore allow precise position recovery. Our fingerprints\´ efficient storage size, extraction time and comparison complexity suggest that our approach can be applied using off-the-shelf PCs and TV "set-top" boxes.