基于三维模型匹配的单张图像物体深度恢复

发布时间：2018-06-08 23:01

本文选题：深度恢复 + 分块匹配　；参考：《浙江大学》2017年硕士论文

【摘要】：从视频和图片中恢复三维信息是计算机视觉领域里的一个基本问题,有着广泛的应用。例如,高质量的深度信息不仅可以直接用来重建场景的三维结构、辅助机器人导航,而且还能帮助计算机更好地识别和理解图像中的物体和场景。特别是随着机器人、虚拟现实和增强现实等应用的兴起,如何让计算机更好地根据所拍摄的图像或视频去分析和理解三维世界变得越来越重要。由于很多时候,我们只能获取单张图像,因此如何为单帧图像恢复深度具有重要的理论研究意义和应用价值。随着三维模型数据集的快速增长,许多研究人员开始尝试利用这些先验模型来进行单帧图像的物体深度恢复。但是,现有的方法一般要求输入图片与训练集中的数据有较高的匹配度,难以处理输入图片中的物体与模型库中的三维模型差异较大的情况;而且,很多方法要求物体的分割信息作为物体深度恢复的先验信息,这很大程度上限制了算法的应用场景。针对这两个问题,本文提出了一个无需预先分割的单张图像物体自动深度恢复方法。首先,将图片与训练集中的数据进行图片匹配,求得候选集,并根据候选集图片的结构特征进行自适应分块,在此基础上进行分块匹配,将分块匹配的结果作为深度预测的依据,有效解决了输入图片与数据集中的三维模型无法很好匹配的情况。其次,基于分块匹配的结果对物体进行自动分割,实现了高质量的物体自动分割和深度估计。再次,针对通过匹配得到的深度在物体不连续边界区域比较粗糙的情况,提出使用改进的分割结果来进行约束,对空洞区域进行深度补全,并进行双边滤波使得深度平滑化,从而有效提高了深度恢复的质量。一系列的实验结果和与其它方法的对比证明了所提方法的有效性。
[Abstract]:It is a basic problem in the field of computer vision to recover 3D information from video and picture, and it has been widely used. For example, high quality depth information can not only be directly used to reconstruct the 3D structure of the scene and assist the robot navigation, but also help the computer to better recognize and understand the objects and scenes in the image. Especially with the rise of robots, virtual reality and augmented reality, it becomes more and more important for computer to analyze and understand 3D world according to the images and videos taken. Because we can only obtain single image in many cases, how to restore depth for a single frame image has important theoretical research significance and application value. With the rapid growth of 3D model data sets, many researchers have begun to use these priori models to restore the depth of objects in a single frame image. However, the existing methods generally require a high matching degree between the input image and the data in the training set, so it is difficult to deal with the difference between the object in the input picture and the 3D model in the model base. Many methods require object segmentation information as the prior information of object depth recovery, which limits the application of the algorithm to a great extent. In order to solve these two problems, an automatic depth recovery method for single image objects is proposed. Firstly, the images are matched with the data of the training set, and the candidate sets are obtained. Then, according to the structural features of the candidate sets, the images are self-adaptively partitioned, and then the blocks are matched. The result of block matching is taken as the basis of depth prediction, which effectively solves the problem that the 3D model of input image and data set can not match well. Secondly, based on the result of block matching, the automatic segmentation and depth estimation of high quality objects are realized. Thirdly, in view of the rough depth obtained by matching in the discontinuous boundary region, an improved segmentation result is proposed to constrain the cavity region, and the depth is smoothed by bilateral filtering. Thus, the quality of depth recovery is improved effectively. A series of experimental results and comparisons with other methods show that the proposed method is effective.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】