arbitrary calibration intrinsic matrix

[tobycollins]

Thanks for this great work. Just one important question. I would like to try this with images using an intrinsic matrix, let's call this K. I see that the network was trained using a fixed intrinsic matrix, let's call it Khat.

The way to do this would seem to be to warp my images by a corrective homography H = Khat*inv(K). Then I can pass the warped images to the network. This seems fine, but the main problem with this is when K has a longer focal length than Khat, my warped image will need to be padded at the borders. Standard padding techniques (zero padding) gives very bad results.

Can you see any simple way that this can be resolved? It's related to you mentioning cropping as a way to use a different intrinsic matrix, however if the focal length is longer we can't deal with it by cropping.

Much appreciated for your great paper and code, Toby

[benjaminum]

We are aware of this problem but unfortunately don't have a solution for this yet.
In this case it is better to not adjust the intrinsics, although the network was not trained with the parameters.

We are happy to hear that you like our work, thanks!