| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 96 === The rapid advance on modern technology has significantly enhanced the procedure of video capture, both in quality and performance, thus the inundation of video data. Video abstraction or video summarization, which focuses on how to quickly and efficiently extract useful information from the raw and huge amount of video data, has therefore become a very important research issue in recent years. Numerous approaches have been developed but each has its drawbacks or limitations.
Inspired by the work of Avidan et al. on seam carving for content-aware image resizing, Chen et al. generalize the idea to perform video carving by repeatedly finding and removing the 2D sheet with the minimal cost from the video volume, thus achieving the temporal reduction of a video. However, there are several issues remained in Chen et al.'s approach, which leaves room for further improvement.
First, the extracted 2D sheet may not be smooth, that is, two spatially adjacent pixels may sit on two different frames that are temporally adjacent, thus causing periodically fragmented results. Second, the employed graph-cut algorithm entails a lengthy computation time. Finally, the required memory is relatively large, and as a result, the input video has to be broken into subsets so that each subset could be brought into the memory in its entirety and processed accordingly. Worse yet, such a partitioning scheme may lead to undesired sub-optimal results.
The main contribution of this paper is to propose a new approach that addresses all the mentioned issues in one shot. First, by modifying the dynamic programming approach originally adopted in Avidan et al.'s seam carving work, we could extract smooth 2D sheets and thus avoid fragmented outlook. Second, unlike Chen et al.'s graph-cut algorithm, our approach is much simpler and efficient, and could perform similar tasks but with a speed that is about two orders of magnitude faster. Finally, the memory consumption is also greatly reduced to be one order of magnitude smaller.
To make our work more complete, we also propose methods to extend the length of a video, as well as to resize a video in spatial domain. Results are shown and compared with existing approaches, if applicable, to demonstrate the effectiveness of our proposed approach.
|
|---|
