I guess that nowadays CV is actually two fields that kinda share the same name. One would be neural-network based and the other one the good old-fashioned CV.

I guess that nowadays CV is actually two fields that kinda share the same name. One would be neural-network based and the other one the good old-fashioned CV. The answer depends on what kind of CV you're looking for. As for the old-fashioned CV I think that visual odometry, bundle adjustment and dense reconstruction are great things to play with and they have practical applications. One hard-ish problem I used to work at is city-scale 3D reconstruction from photos. You can have street-level image data you collected to get the high detail and some aerial pictures to get the consistency for wide-area. A problem of matching street-level and aerial pictures is even interesting from research point of view (keyword is wide baseline image matching). With this you can build something similar to what Google Map has where you point your phone on the street and you get overlay directions. And if you call it mixed-reality metaverse, you can start a startup 6 months ago. As for the NN-based I'll leave that for others.