I am implementing an algorithm to remove projective distortions on the following image.

The image shows two irregular shapes which will be subject to projective distortions

I understand this is possible by applying the following transformation:

$$ \begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ l_1 & l_2 & l_3 \\ \end{matrix} $$

Where $$ l_\infty=(\begin{matrix}l_1 & l_2 & l_3 \end{matrix})^T $$ is the line at the infinity. In a perspective image of a plane, the line at infinity on the world plane is imaged as the vanishing line of the plane.

The vanishing line can be computed by intersecting two vanishing points which can be computed in the following ways:

  1. From the intersection of two sets of imaged parallel lines. But seems there are no two sets of imaged parallel lines on the image.
  2. Given two intervals on a imaged line $$\lt0,a^\prime,a^\prime+b^\prime\gt$$ with a known length ratio $$ d(0,a^\prime):d(a^\prime,a^\prime+b^\prime)=0:a^\prime $$ Where I need to solve the system (up to scale) $$\left(\begin{matrix}0 \\ 1\end{matrix}\right)=\left(\begin{matrix}h11 & h12 \\ h21 & h22\end{matrix}\right) \left(\begin{matrix}0 \\ 1\end{matrix}\right)$$ $$\left(\begin{matrix}a \\ 1\end{matrix}\right)=\left(\begin{matrix}h11 & h12 \\ h21 & h22\end{matrix}\right) \left(\begin{matrix}a^\prime \\ 1\end{matrix}\right)$$ $$\left(\begin{matrix}a+b \\ 1\end{matrix}\right)=\left(\begin{matrix}h11 & h12 \\ h21 & h22\end{matrix}\right) \left(\begin{matrix}a^\prime+b^\prime \\ 1\end{matrix}\right)$$ And compute the vanishing point as $$x^\prime=\left(\begin{matrix}h11 & h12 \\ h21 & h22\end{matrix}\right) \left(\begin{matrix}0 \\ 1\end{matrix}\right)$$ But I don't understand this approach as I don't have the world points $$\lt0,a,a+b\gt$$ and I don't know what serves me for knowing the length ratio.
  3. Using the cross ratio. Which I totally don't understand how could be possible to use in this case.

I would appreciate any insight about this.

Edit: isn't necessary to follow any particular approach just remind that the planar objects are irregular (no orthogonal angles in real world) congruent shapes (they have the same shape in real world)

  • $\begingroup$ I have written an answer on SO about how to compute a projective transform given four points and their images. You can use vanishing points for my approach, but don't have to. $\endgroup$ – MvG Feb 4 '13 at 15:36
  • $\begingroup$ Thanks for your answer, well actually using directly a projective approach based on four points is an oversepecification of the geometry (8 degrees of freedom instead of 4). Furthermore i don't count with such 4 pairs of points for which i know its shape in world frame; the piece of paper under the rightmost pen might not be rectangular or even may not exist. $\endgroup$ – Andres Felipe Feb 5 '13 at 12:05
  • 1
    $\begingroup$ I see I haven't read your question carefully enough, but now I have a proper solution for you, posted as an answer. If that answer is what you are looking for, then perhaps you should modify the title of this question, since you don't necessarily need to use cross ratios for this. Nor some real vanishing points, come to think of it. So “Reconstruct perspective from congruent shapes” or something like this would better describe the task you have at hand. Except you really want a solution based on cross ratios, that is… $\endgroup$ – MvG Feb 5 '13 at 14:18

I will assume that the two depicted bright polygons are congruent in the world plane. If they are not, then I guess you won't have enough information to reconstruct anything.

The first step is finding a projective transformation which maps one of the polygons onto the other. That transformation is uniquely defined using four points and their images, and I've described this computation in detail in another post. So choosing any four pairs of matching corners will give you that matrix. Originally I had assumed that this is the transformation you were interested in, but I had not read the question carefully enough. The transformation you just found describes the rotation of one polygon onto the other in the image plane.

Next, you look for fixed points of the transformation. You find these as eigenvectors of the transformation matrix. In $\mathbb C\mathrm P^2$ you should get three fixed points, one of them real and two complex and conjugate to one another. The real one is your center of rotation. The complex fixed points correspond to the ideal circle points $I=(1,i,0)^T$ and $J=(1,-i,0)^T$, which remain fixed under every similarity transformation.

By the way: The join of the two complex points corresponds to the vanishing line. Which in your image is way outside the picture area. But knowing the vanishing line is less useful than knowing the locations of $I$ and $J$ as we do, since these points can be used to define angles, so they will help you avoid skewing.

Now you can again choose four points and their images to define a projective transformation. This time, you map the complex fixed points to $I$ and $J$, and you map the center of rotation to wherever you want it to lie in your reconstructed image, e.g. the origin. You still have one point which you can choose arbitrarily. Its image will fix the scale and orientation of the reconstructed image. Strictly speaking, you don't have to take the center of rotation as one preimage point either: as long as you correctly map $I$ and $J$ from image coordinates to world coordinates, you can choose any two pairs of real points to uniquely define the projective transformation.

Counting degrees of freedom, the two arbitrarily chosen points above amount to four real degrees of freedom, matching the four degrees of freedom of a similarity transformation. Everything else is fixed. In particular, not only parallel lines but also angles are reconstructed.

I've just implemented the above description as a proof of concept using Cinderella. The center of rotation, drawn in green, appears to coincide with the upper boundary of your picture. The points P1, P2 and R2 were chosen arbitrarily.

Original and reconstructed image

  • $\begingroup$ Thanks a lot for your answer its been very helpful. I am writing because I got some issues: 1) When mapping the complex fixed points to the ideal points I am obtaining a complex transformation matrix, is it correct? 2) How should I choose the fourth point? despite the fact of being complex I am applying the projective transformation to the planar objects vertices but it reflects horizontally and vertically. 3) I tried to apply the transformation to the complex fixed points but they didn't transformed onto the ideal points, is it correct? $\endgroup$ – Andres Felipe Feb 6 '13 at 2:48
  • $\begingroup$ @gantzer89: 1) The transformation matrix should be real, as you map conjugate points onto conjugate points. 2) Choose the fourth pair of preimage and image both as real points. Apart from that, tweak it to control the size and rotation of the reconstruction. If you get "two reflections", then you get a rotation, so rotate P2 around R2 (in my example) to align the image any way you want. 3) If the pairs of points defining the transformation don't end up as preimage and image under said transformation, you made a mistake computing its matrix. The complex fixpoints should map to $I$ and $J$. $\endgroup$ – MvG Feb 6 '13 at 3:00
  • $\begingroup$ Thank you for your patience answering my questions, I continue having problems with the projective transformation, in this other question I am showing the matrices I am using for the computation of the transformation. $\endgroup$ – Andres Felipe Feb 6 '13 at 15:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.