【作者】 杨洋；

【导师】 鲍虎军； 章国锋；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2013， 硕士

【摘要】 单目视频的深度恢复是计算机视觉领域的一个很重要的问题,尤其在3D电影越来越受追捧,各种3D设配,3D应用兴起的当下。人们越来越追求3D视觉的享受,通过对单目视频实现深度恢复,可以减少制作3D电影的制作成本,同时可以扩大3D市场的片源。越来越多的研究人员将研究的焦点对准了视频的深度恢复,这不仅仅是商业化市场的需要,同时也是技术发展的要求。在2D转3D的研究中,最近几年的研究热点是如何通过机器学习的方法来实现视频深度恢复,之所以焦点对准机器学习,是因为机器学习的方法,相对于传统视觉几何方法,更具普遍性。传统的通过分析不同视频特点,解决特定问题深度恢复的解决方案,虽然会有更好的视觉效果,可是使用的场景受限,算法的工业化意义降低。本文提出的算法是以视频分割为基础,融入了机器学习的方法,同时为了提高算法的适用性,加入了用户的输入信息。作为输入的用户信息对于恢复视频的深度信息,具有很好的指导意义。本文实现的基于机器学习的单目视频深度恢复,首先考虑到算法的适用性,以及实验重点,所需要的分割信息以及预处理时的深度信息,作为输入信息给定,给定某些帧的分割信息以及深度信息；其次通过GMM模型,基于SIFT流的信息传递以及基于形状先验的分割信息传递,将人工处理的分割信息传递到相邻视频帧,并通过插值法获得前景区域的深度信息；最后对于背景区域的深度恢复,采用SIFT流传递深度信息,采用图像补全技术进行补全。图像补全技术的限制,本文提出采用机器学习的方法,来实现未知区域的深度优化。本文详细的介绍了算法实现的细节,同时实验的结果表明,本文提出的算法具有鲁棒性,以及很好的实用性。更多还原

【Abstract】 Depth recovery of monocular video is an important issue in Computer Vision. Especially,3D movies are more and more popular nowadays and kinds of3D display devices and3D application occupy the market. People are actively going after the3D enjoyment. So the depth recovery of2D video not only reduces the production costs of3D movies, but also gives more visual enjoyment and more3D movies. More and more researchers have focus their concerns on video depth recovery.How to implement depth recovery using machine learning method is a hot topic in recent years. There are lots of researches about this topics having been done and some reasonable algorithms have been proposed. Why concentrated attentions on machine learning are that machine learning method is more general than conventional view geometry method. The results of view geometry method are more accurate and look like much better, but the limitation is that for different input video, we should apply different view geometry approaches using different depth cues.Our method first performs segmentation mask and recoveries the depth map on the key frame by user input, then propagates the segmentation and depth information onto successive frames. We use the inpainting technology to restore the missing pixels’depth, but this method has some limitation in big missing area. So we take machine learning approach to mix up this disadvantage. In our paper, we use SVM method to train learning model and predict the depth information. User’s input will provide lots of guidance information and make the method more general. That’s the reason why we need user’s input.In this paper, we introduce every single implementing detail carefully and the results show that our method outperforms and yields excellent depth maps.更多还原