Frame to frame homography

There are several deep models which estimate the frame-to-frame homography, for example SuperPoint, or SuperGlue.
In the case of finding homography, one assumption is that the camera is a pinhole.

I am using a stationary lens camera (no translation, only ‘rotation’ and ‘zoom’ are the variables). Due to zooming the intrinsics could change. In this case, what is the correct way of finding the homography matrix?


Since you are using a homography, is it safe to assume you don’t have any non-linear lens distortion (or are accounting for it somehow)? If that’s the case, then the focal length change can be modeled by the homography, so no issue there. If you do have lens distortion, the focal length change will affect the distortion model, so you would have to account for that before trying to use a homography to map between two images.

Thanks a lot Steve!
I need to check if the distortion is non-linear.