> Once our geometry reconstruction has high accuracy for hard-surface objects we'll move on to this category, collect data, train and improve.

is that really possible though? For complex objects there should be multitude of 3d structure that fit a 2d projection and there's really no way to say "which one is the most correct one".