letatanu / deephomography

The unofficial implementation of the paper Deep Image Homography Estimation.

Deep Image Homography Estimation

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.

Dataset

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.

Pre-processing the dataset.

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.

Training

The model was trained over 49 epochs. The initial learning rate is 0.005, and it is divided by 10 per 30000 iterations.

Result

The average corner error after 49 epochs is:

1. In the train set: 6.003
2. In the validation set: 6.034

Demo

You can run the Demo.py with the pre-trained model to see the predicted reusult.

python Demo.py --image <image url>

deephomography's Issues

Doesn't work for identity scenario.

pakage version

Can you tell me the version of python and all the package version?
I have a problem with loading model, maybe it's caused by the wrong version.
thank you.

Demo doesn't work - pre trained model missing in the repo.

I don't think you have pushed the pre-trained model in the repo. That's the reason your demo doesn't work.

Doesn't work when same patch is fed as inputs.

If I feed the same patch (without any perturbation);
The target will be [[0,0],[0,0],[0,0],[0,0]] in H_4point format or 3x3 identity matrix in H_matrix form. But the output from the model is pretty random. It doesn't work in this case.
Any thoughts on why this is the case?

It also doesn't work in the case of simple translation in the x-direction.
The target will be [[16,0],[16,0],[16,0],[16,0]] in H_4point format.