I'm still not sure how this was done. Here's what I have gathered so far from the post: The author has two images taken unknown location. He scaled them and then marked points on each. Then he matched up points on two images manually. Now he has dx and dy for each point on one image, relative to other. Now he asserts that bigger dx means nearer to camera. So I'm thinking he takes some proportionality constant to get z = c * dx.

But wouldn't that produce pretty arbitrary shape depending on value of c?