“Including all the noise …” That’s a great insight.
So ‘starting afresh with depth maps from the aligned photos’ is just that - a fresh start (for better or worse) from whatever quality of photo alignment has been achieved (could be great but could also be error-full) but at least it’s without the ‘noise’ accumulated in the ‘try this/try that’ of getting there.
Personally I still don’t understand how a data point in a depth map is any different from a point in a point cloud, nor how that data point is calculated if not by projection to tie point (just like the point cloud).
But they say it’s not only superior, but v much quicker than computing a dense point cloud - hence RC’s speed?