Semiautomatic scene modeling from a
single view of partially known structure

I found one of Paul Haeberli's images.

The steps needed to calibrate camera and extract scene geometry.

  1. detect image lines.
  2. detect at least two image points.
  3. group parallel lines.
  4. group coplanar lines and points.
  5. define world relative directions of 3 orthogonal edges.
  6. define world relative positions of 2 point features.
Here's the resulting camera and scene geometry wire-frame rendered from a few different views.
Copyright © 1995 by MIT Media Lab. All rights reserved.