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
| dataset | image numbers |
|---|---|
| train2014 | 82783 |
| val2014 | 40504 |
| test2014 | 40775 |
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:
