1
$\begingroup$

As I was searching for available resources( here, here, here and many more), I realize that many developed optical algorithm is to estimates LARGE displacement. However, there is very limited information or journals that put effort on estimate very small displacement. The small displacement I'm taking here is at sub-pixel level. The smaller the resolution the better.

So my question, you guys have any idea in mind that might inspire a new method, or suggest an existed method (that I might not be aware of/that can be improved)? Any method/suggestion/comment is very welcomed!

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

I was confronted to the same problem and I can recommend to use particle filtering. The idea is that instead of using a discretization of the space (which would favor only a limited range of motion detection), the optic flow is computed as a probability density function which is represented by a limited number of samples. These samples (or particles) are precise up to the precision of your computer and can thus detect very small displacments.

Some resources:

Improve this answer
$\endgroup$
4
  • $\begingroup$ This is a very helpful and informative answer! I don't even know the existence of this kind of filter! Thank you so so much! Now I can progress from there. I see you're the author of the marvelous work above, I would like to take this chance to congratulate you on the amazing work! Is it okay if I refer back to you here in case I encounters problems after going deep into your work? $\endgroup$
    – Gregor Isack
    Commented Apr 8, 2017 at 3:10
  • 1
    $\begingroup$ thanks for the feedback - all the code is available on the git repo - do not hesitate to contact me directly $\endgroup$
    – meduz
    Commented Apr 14, 2017 at 8:41
  • $\begingroup$ Hi meduz, I got a problem while interpreting your code. Hope you could help me out! In your code line 1020, is that the intensity of the image at location x_px and y_px? Since it's an integer, wouldn't the likelihood will be very likely the same for all the generated particles? And is the x and y defined in your code is the pixel location on the images, respective to x-axis and y-axis, or it's pixel displacement?Thanks! $\endgroup$
    – Gregor Isack
    Commented Apr 19, 2017 at 6:02
  • $\begingroup$ you are right, the likelihood is computed on the data that is available, that is, on the pixel values from the input movie (but you could think about other input representations...). Still your belief (pdf on motion) can be sub-pixel because it is represented by the particles. Do not hesitate to create an issue using github's tracker if something appears unclear. $\endgroup$
    – meduz
    Commented Apr 20, 2017 at 10:52

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.