Multi-video Object Selection Based On Co-segmentation

With the popularity of smartphones, tablet computers and other electronic products, as well as the rapid development of media platforms, such as WeChat and MicroBlog, image and video have deeply affected people's way of life. As one of the most abundant information carriers, video has a broad application in various industries. With the explosive growth of video information, how to make the computer understand the content of the video scene, and enable people to obtain the required information quickly and efficiently are becoming more and more important. Video segmentation as the basic step of video content analysis, plays a key role for post-processing of video information, therefore also has been more and more attention from researchers. In addition, as a more challenging task, video co-segmentation uses the characteristics consistency among multi-videos to co-segment the common target of the video set, which greatly improves the single video segmentation efficiency.The article mainly studied some problems in the field of video object segmentation, took appropriate co-segmentation model to establish the relationship between multi-videos, and solved the irrelevant frame(frame containing no targets) interference problems, as well as by acquiring object tracklet with global consistency constraint to provide reliable guidance for co-segmenting multi-videos. The contributions of this paper can be concluded as follows:Firstly, this paper introduced the key technologies of video segmentation involved in, including object proposal generation methods and saliency detection methods, and then compared the effectiveness of several different methods, as well as introduced common targets multi-search strategy and the motion information detection method based on optical flow, which provided theoretical basis for the new co-segmentation methods.Secondly, this paper proposed a novel multi-video object co-segmentation method to co-segment the videos which contain irrelevant frames(frames containing no common targets). At first, we generated a variety of object proposal candidates and grouped them into multiple target clusters. Then, we constructed a weighted graph based on clusters and selected the common object clusters. Based on the selected clusters, we proposed a discriminant mechanism to detect irrelevant frames, and used a graph-based framework to refine pixel-level segmentation of relevant frames.Thirdly, this paper presented a video co-segmentation algorithm using object tracklet. By using the relationship of adjacent frames and global consistency, we tracked the target proposals and obtained the object tracklet, so as to get reliable target motion trajectory and priori information, and finally completed segmentation results by optimizing graph-cut energy function.