Dmytro Mishkin

I'll merge, but let's add that to the main as well

@yanirmr there is no bug, but it seems that the kornia conventions are not clear. The output similarity matrix is normalized, so the output is not in pixels, but pixels/imgsize...

@yanirmr the reason we use not pixels, is that such homography is the same, regardless the image downscale or upscale. That's useful for image pyramid in particular.

KeyNet detector detects variable number of keypoints per image, same for the GFTT, so it is practically impossible to do the batch processing there. The documentation needs to be fixed,...

Yes, I can take care of it

@zechengtang could you please add the test cases?

@georg-bn please do!

I believe it is not...

@007anupamdas will fix, thank you