This kind of thing is where having access to extra information can definitely provide some benefits. I’ve previously extracted video frames by choosing the sharpest frame within a short time period, at regular intervals (so lower chance of bad photos because you’re choosing the best one around each point in time), but that could definitely be improved by aligning the telemetry with the recorded video, and taking account of integrated distance from the accelerometers, as well as things like vibration and speeds to determine the best frames to extract.
Come to think of it, that could make for a decent introductory (ish) programmatic log file analysis tutorial or something. I’ll add it to my list 
For photogrammetry purposes it’s of course preferable to use raw images if possible, rather than ones that have already been compressed for video, but that’s generally only possible with some more specialised hardware. I’ll have to see if I can work something like that into StaROV, since I’m already planning to use an OAK camera in it. Hopefully I can set it up to send raw frames on command based on telemetry and timing, in a separate queue to the normal encoded video stream.