This project is the unofficial implementation of the paper Deep Image Homography Estimation.
A homography is a mapping from a projective space (image) P to Q. From this network, it will estimate
a 4-point homography parameterization which maps the four corners from one image into the second image.
| An example | Result | Patch | GroundTr | Predicted |
|---|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
The pre-trained model is uploaded here.
I used MS-COCO dataset as described in the paper. You can download it from here. There are 118287 images in the train set, and 40670 in the test set.
Resize all images in the train and validation sets to the size of 320x240 and resize all images in the test to the size of 640x480. Because I wanted to speed up the train progress, I then combined the resized images in train, validation and test sets into HDF5 files. For more detail, please take a look at the Dataset.py in the models directory.
The model was trained over 49 epochs.
The initial learning rate is 0.005, and it is divided by 10 per 30000 iterations.
The average corner error after 49 epochs is:
- In the train set: 6.003
- In the validation set: 6.034
You can run the Demo.py with the pre-trained model to see the predicted reusult.
python Demo.py --image <image url>