0

I have two images which are identical except that one of them is slightly distorted (e.g., the image is stretched in the middle.)

I would like to define a fine grid of points on the original image and track their position on the distorted image. Note that the tracking points are arbitrary.

Could anyone please help me find an algorithm that can handle this. I am very new to this field so any elaboration is much appreciated.

The following images are an example for this question.

Original Image:

img

Distorted Image:

img

Thanks!

Spektre
  • 49,595
  • 11
  • 110
  • 380
Matt Sog
  • 1
  • Manual or automated ? And what do you mean by "the tracking points are arbitrary" ? –  Sep 21 '17 at 07:21
  • SIFT/SURF+RANSAC + polynomial curve fit and interpolation of the grid mapping something like linearization in here [OpenCV Birdseye view without loss of data](https://stackoverflow.com/a/39316776/2521214) – Spektre Sep 21 '17 at 07:33

2 Answers2

0

Check out these tools. They both of the ability to automatically or manually add tracking points between multiple images and show the difference.

http://hugin.sourceforge.net/

https://www.ptgui.com/

pucky124
  • 1,489
  • 12
  • 24
0

I see four steps in this process:

  • finding the tracking points on both images. You can do it manually, or using a so-called interest-point detector.

  • matching the corresponding points in both images. Again, you can do it manually of using a so-called interest-point descriptor and a matching algorithm.

  • fitting a deformation model to the point pairs. Many options are possible such as bivariate polynomial or bivariate cubic splines. (You can also think of triangulating the points but the degree of continuity will be poor and artifacts noticeable).

  • warping the deformed image with this model.

None of this is elementary and there are many possible combinations. I doubt that you will find a ready-made solution. But you can get some inspiration from image stiching software, which use these four steps, with a specific, simple deformation model (homography).

I expect the automatic matching approach to be unreliable due to the symmetry of the shape that implies numerous similar points, causing ambiguity.