Deep Image Homography Estimation

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:

 Deep Image Homography Estimation

 

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:

Deep Image Homography Estimation

 

训练方式:SGD(随机梯度下降法) ,momentum =  0.9

训练数据制作:

Deep Image Homography Estimation

训练标签:

Deep Image Homography Estimation,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