1
$\begingroup$

Assuming I have two sensor feeds, is there an algorithm that gives an estimate of how likely it is that they point at/derive data from the same scene (possibly from different angles, with different modalities, filters and perspectives)?

Perhaps a measure of how easily the data streams could be related/reduced to a common cause?

Simple correlation does not seem to suffice here, since the signals may only be indirectly (and nonlinearly) related and do not necessarily have the same "shape".

This seems to be related, but not the same as the image registration problem.

Improve this question
$\endgroup$
4
  • $\begingroup$ Can you limit the use case more? Be more specific? $\endgroup$
    – Royi
    Apr 11 at 12:40
  • $\begingroup$ Could you please review my answer? $\endgroup$
    – Royi
    May 22 at 9:48
  • $\begingroup$ @Royi Someone edited my answer to be about images, which it wasn't specifically, so I'm looking for something more general $\endgroup$
    – 2080
    May 22 at 17:13
  • $\begingroup$ Can you describe your data? Because even the original question seems to imply on images (Scene, angles, perspective). $\endgroup$
    – Royi
    May 22 at 17:15

3 Answers 3

Reset to default
1
$\begingroup$

When you say "same shape" and "coincidence" I think at symbolic meaning. I'm not sure this is what you want but if it is so you can use CNN to extract symbolic features from different noised data.

Do you want to recognize the same object from different images, and to extract the key to recognize them in every different situation? Compression is a different task.

Improve this answer
$\endgroup$
0
+25
$\begingroup$

The current formulation of the question is pretty broad.

Basically what you're after it metric learning, you have 2 signals and you ask what's the correct metric to say if they close or not.
This is the general question which is really hard to do.

If you have a data base of many images which are similar to your case and you can label them, you may create a model as a classifier.

If you can't label data you have to use classic computer vision.
I'd go with a pipeline as following:

  1. Color Balance
    We need both images to have the same white / color balance.
  2. Histogram Comparison
    Sanity check by having somewhat similar colors distribution. You may use one of many metrics for distributions.
  3. Object Detection
    Use pre defined model to compare object on both images.
  4. Image Registration
    Given all above work you may try to register the images. If it works with high quality you may infer the images are looking at the same scene.
Improve this answer
$\endgroup$
-1
$\begingroup$

The classic way to do is to detect image features and use descriptors to compare images but its performance will depend upon how robust your features are to different transformations..

Improve this answer
$\endgroup$

Your Answer

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

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