单幅图像深度重建SFT.ppt

4.基于纹理变化的深度估计算法 SFT的传统方法是通过建立一个随机场模型描述图像中纹理信息，通过学习得到模型参数，实现场景深度重建，但由于此类方法使用的是局部的、二维纹理信息，因此所建立的模型中无法完全反映出场景中各目标间的空间关系。近几年来，针对此类问题的研究逐渐向使用全局的、多尺度的、分层次的纹理信息方向发展，并已经开始应用于实际系统中。 2006年，斯坦福大学的Andrew Ng等人应用机器学习的方法，采用分层的多尺度马尔科夫随机场模型，通过对大量的训练集数据的学习，建立室外场景图像中各目标间的空间关系模型，并将取得的结果应用在无人汽车的自主导航中。 这种利用全局的三维几何纹理信息的方法，实际上就是在利用了图像“质地变化率”的同时也利用了图像中的遮挡信息（空间关系）。 3-D Depth Reconstruction from a Single Still Image Ashutosh Saxena, Sung H. Chung, Andrew Y. Ng ,2007 4.1 Visual Cues for Depth Perception Monocular Cues texture variations, texture gradients, interposition, occlusion, known object sizes, light and shading, haze, defocus, etc. global information, local information. Stereo Cues Motion Parallax and Focus Cues 4.2 Feature Vector The image is divided into small rectangular patches, and estimate a single depth value for each patch. Two types of features: absolute depth features—used to estimate the absolute depth at a particular patch—and relative features, used to estimate relative depths. Feature Vector three types of local cues: texture variations, texture gradients, color. Capture cues Laws’ masks(3*3)------- texture variations Oriented edge filters(6)------- texture variations Color channel(2)------- color 2.1 Features for absolute depth Features for absolute depth initial feature vector---sum absolute energy and sum squared energy, multiple spatial scales---capture more global properties of the image, immediate neighbors---capture occlusion relationships, vertically column---show vertical structure. absolute depth feature vector x is 19 ? 34 = 646 dimensional. 2.2 Features for relative depth A different feature vector is used to learn the dependencies between two neighboring patches. a 10-bin histogram of each of the 17 filter outputs |I ?Fn|, giving us a total of 170 features yis for each patch i at scale s. relative depth features yijs for two neighboring patches i and j at scale s will be the differences between their histograms. 4.3 Probabilistic Mode