Hi Yegor
all your mentioned stuff is already implemented
Looking on the dataset you’ve sent to me, I can see one primary problem which can be solved pretty easily… Just take MORE images. you have 40 images altogether in 3 elevations and it’s NOT enough.
It‘s recommended to take one image every 10 degree per elevation (36 images - 18 alternatively good too, if more elevations are present) and for the subject like you try to do, I would say at least 4-5 elevations - 144+ images. and you’ll always get perfect results.
So in short, the Tx reprojection issues can be traced to bad alignment as there is not enough overlap among images.