Deep Image Homography Estimation理解_综合

文献：Deep Image Homography Estimation，下载地址

输入：128x128x2

Padding：'SAME'

池化步长：2

回归模型（HomographyNet-Regression）：

conv1 3x3 : 128x128x64

conv2 3x3 : 128x128x64

maxpooling1 2x2: 64x64x64

conv3 3x3 : 64x64x64

conv4 3x3 : 64x64x64

maxpooling2 2x2: 32x32x64

conv5 3x3 : 32x32x128

conv6 3x3 : 32x32x128

maxpooling3 2x2: 16x16x128

conv7 3x3 : 16x16x128

conv8 3x3 : 16x16x128

fully connect1: 1024x1

fully connect2: 8x1

loss function:

$\frac{1}{2}*\left \| p\left ( x \right ) -q\left ( x \right )\right \|^2$

分类模型（HomographyNet-Classification）：

conv1 3x3 : 128x128x64

conv2 3x3 : 128x128x64

maxpooling1 2x2: 64x64x64

conv3 3x3 : 64x64x64

conv4 3x3 : 64x64x64

maxpooling2 2x2: 32x32x64

conv5 3x3 : 32x32x128

conv6 3x3 : 32x32x128

maxpooling3 2x2: 16x16x128

conv7 3x3 : 16x16x128

conv8 3x3 : 16x16x128

fully connect1: 1024x1

fully connect2: 8x21

softmax

loss function:

$-\sum p\left ( x \right )log\left ( q\left ( x \right ) \right )$

训练方式：SGD(随机梯度下降法) ，momentum = 0.9

训练数据制作：

训练标签:

$\left [ \Delta u1,\Delta v1,\Delta u2,\Delta v2,\Delta u3,\Delta v3,\Delta u4,\Delta v4 \right ]$ ,与放射矩阵H一一对应

训练设置：conv8与fully connect1需要添加dropout=0.5

测试数据集：MS-COCO

运行效率：NVIDIA Titan X GPU， 300fps