1
$\begingroup$

Coming from a photogrammetry background, I am quite familiar with camera calibration, and the common assumptions that are used for crowd-sourced images (i.e. center point in the middle of the image, pixel resolution from the EXIF data etc.).

But I have been wondering a maybe naive question lately: Can I calibrate a manipulated image (e.g. downloaded from twitter or facebook), that has possibly been cropped or maybe even resampled and has no EXIF data? The nice thing is: More than 4 ground control points in the desired object coordinate system can be asummed available.

Is there any papers available on this problem?

Improve this question
$\endgroup$
2
  • $\begingroup$ Updated my reply. $\endgroup$
    – Tolga Birdal
    Commented May 26, 2014 at 8:26
  • $\begingroup$ Why wouldn't you try 4 point minimal pose and unknown focal length problem, as I mentioned? $\endgroup$
    – Tolga Birdal
    Commented Jun 20, 2015 at 20:35

1 Answer 1

Reset to default
1
$\begingroup$

1) It seems to me that you don't care about solving for correspondences or obtaining any high level paradigm in that sense. So I would treat this problem as a minimal problem and redirect you to check:

http://cmp.felk.cvut.cz/minimal/

Maybe specifically, 4-point method with unknown focal length: http://cmp.felk.cvut.cz/minimal/p4pfr.php

You could find many variants in that page.

2) Well, this problem involves many issues of multiple view geometry theory. However, in this case, an uncalibrated approach would be more suitable.

You could obtain point correspondences from SIFT-like features. As your images are downloaded from social platforms, it is likely that they will contain a lot of such feature points. Then, you could match the descriptors in a RANSAC-like fashion and retain the correct matches. From there it is possible to compute $F$ the fundamental matrix. With further assumptions, constraints or knowledge about the scene you could compute the scale as well as the 3D point coordintates. Some examples here:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.31.4741&rep=rep1&type=pdf

Under certain constraints it is also possible to recover the camera matrices:

http://users.cecs.anu.edu.au/~hartley/Papers/eccv92/Higgins/higgins.pdf

After all, I would recommend you to read:

http://www.robots.ox.ac.uk/~vgg/hzbook/

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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