This is an implementation of the paper Deep Image Homography Estimation with tensorflow
网络参数下载地址 百度网盘

ms-coco

Build the network according to the paper completely.
No need to pre-save the generated data, the program genetated the image pairs automatically.

Test only one image a time, output the four-pair offsets predicted from HomographyNet.
Then use data_process.m to visulize the results.

Use existing cnn framework for training, something wrong v.

If you like, just generate the training data.

I test 200 images on test2014, Mean Corner Error = 12.6578 (image size is 320x240).
The original thesis is 9.2 (image size is 640x480). But I believe my result could be better.
good example:

my result Vs. the author's result on the same image

bad example: