Partial image reconition

We have a challenge detecting an image larger than the camera frame.

We have developed image recognition, where we detect an image (target model) in a video stream, calculate the homography and then the 3D pose.
This works nice and smoothly.

We use AKAZE to extract features from the taget model, and compare these feature with the one we extract from the camera feed. If we have a match we calculate the homography.

Now we got the challenge to do image recognition on a very large image. The camera feed is only able to see part of the image - roughly 20%. We are not able to calculate a valid homography.

Is it possible to calculate a valid homography from a part of a larger image model, or do we need to slit the image up in several minor model targets?