Hey guys,
nice work you have done here! I was able to speed up my orb slam by factor of 4. However, for initialization somehow only a small part of the image is used for matching (see attachment). As soon as the orb slam is initialized, features are tracked over the image as a whole, but before that only that small area is used. This gets even worse, if I use bigger images.
Any suggestions which part of the code might do this?

Thanks and best regards,
Alex