解读Depth Map Prediction from a Single Image using a Multi-Scale Deep Network （1）

解读Depth Map Prediction from a Single Image using a Multi-Scale Deep Network （1）

我不会把这些翻译成中文，一是麻烦，二是因为英语是无法回避的！

先来看一下摘要，可以得到一些信息：

(1) Predicting depth is an essential component in understanding the 3D geometry of a scene

(2) A new method is presented to find depth relations from a single view by employing two

     deep network stacks:

    a. makes a coarse global prediction based on the entire image

    b. refines this prediction locally

(3) apply a scale-invariant error to help measure depth relations

再来看看Inroduction和Related work,可以得到前人在做深度估计的成就

(1)Provided accurate image correspondences, depth can be recovered deterministically in the stereo

     case [5]

(2) predict depth from a set of image features using linear regression and a MRF [15]

(3) integrate semantic object labels with monocular depth features to improve performance [12]

(4) use a KNN transfer mechanism based on SIFT Flow to estimate depths of static backgrounds

      from single images, which they augment with motion information to better estimate moving

       foreground subjects in video [7, 11]

当然，我的摘抄分析不一定完全。从他人成就来看，用CNN处理单目深度估计的思想是超前的

来看看他们怎么叙述自己的Approach，

从图1（Model architecture），很容易了解CNN网络处理图片的过程

如果你像我一样不明把conv,stride,pool的含义，可以参考这篇博文点击打开链接

学习这篇博文中。。。。。。