解读Depth Map Prediction from a Single Image using a Multi-Scale Deep Network （7）

解读Depth Map Prediction from a Single Image using a Multi-Scale Deep Network（7）

上次分析了CNN网络的理论，以及讨论了该文章的CNN设计思路以及误差函数的选择，以及对输入数据做

变换操作确保算法训练具有鲁棒型，

今天说一下实验部分，

可能读者会抱有疑问，文章的理论部分懂了，那么剩下的实验部分有必要了解吗？这要因文章而定

NYU depth数据集的特点：

1, composed of 464 indoor scenes

2, taken as video sequences using a Microsoft Kinect camera

用数据集训练的方式：

1, using 249 scenes for training and 215 for testing

实验对数据集的处理：

1, RGB inputs are down-sampleed by half, from 640*480 to 320*240

2, Because the depth and RGB cameras operate at different variable frames rates, we associate each depth

     images with its closest RGB images in time,

     and throw away frames where one RGB images is associated with more than one depth

3, we use the camera projections provided with the dataset to align RGB and depth pairs

     pixels with no depth value are left missing and are masked out.

4, to remove many invalid regions caused by windows, open doorways and specular surfaces we also

      mask out depths equal to minimum or maximum recorded for each image

可见，即便理论是完善的，在实际的处理中，还是要对数据集考虑很多问题，以确保训练的正确性。

训练的具体过程：

1, The training set has 120K unique images

2, We test on the 649-image NYU Depth v2 test set (with filled-in depth values)

3, We train the coarse network for 2M samples using SGD with batches of size 32.

     随机梯度下降法（Stochastic Gradient Descent，简称SGD

4, We then hold it fixed and train the fine network for 1.5M samples (given outputs from the already-

     trained coarse one)

 

实验时间：

Training took 38h for the coarse network and 26h for fin, for total of 2.6 days using a NVidia GTX Titan Black

通过对一个数据集非常完整的训练过程，可以了解到一些非常具体琐碎但又非常重要的一些事情！

下一次来分析相应的工程Python代码