0
$\begingroup$

When using an RGB camera, what is the minimum number of pixels of an object that you need to identify something (theoretical or rule of thumb)?

The image would be taken out in the wild with a complex scene and the objects would not be known to be there but they will be in a list of known candidates.

By identify I mean, from an unknown object can you detect it and discern whether it is a tree, a monkey or a sheep as opposed to just seeing a blob and only being able to tell it is a blob.

$\endgroup$
5
  • 3
    $\begingroup$ one, if you are in a desert and there is only one tree $\endgroup$
    – jsotola
    Commented Jul 26, 2021 at 2:32
  • $\begingroup$ except that you wouldn't know if you were seeing a tree or sensor noise - assume you only have one chance to take a single image. And even if you could take repeated images then you might deduce there was something there but you still would not be able to tell if it is a tree or a building or anything else that could be a similar colour to a tree... $\endgroup$
    – qwertyuiop
    Commented Jul 26, 2021 at 5:36
  • $\begingroup$ your post said nothing about taking a single image ... also, your question is opinion based ... different people require different amounts of visual information in order to positively identify an object $\endgroup$
    – jsotola
    Commented Jul 26, 2021 at 5:41
  • $\begingroup$ you are correct I didn't, but I also did not say I would take multiple images either. Also I do not think that anyone would be able to successfully id any thing based on a single pixel even in the contrived situation you have dreamed up. $\endgroup$
    – qwertyuiop
    Commented Jul 26, 2021 at 6:07
  • $\begingroup$ you could use one pixel sensor and mechanically scan the scene ... something like an early form of TV or a laser barcode scanner $\endgroup$
    – jsotola
    Commented Jul 26, 2021 at 16:27

2 Answers 2

2
$\begingroup$

There is no really precise answer on your question. There are a lot of parameter would affect number of required pixel, what kind of detection method you use, how do you preprocess the image, noise of image and etc.

Say you perform color based detection. You wanna if there is object colored red somewhere on your image. Then how many pixel needed? One.

If you perform marker based detection, then depends on the marker itself and algorithm to detect it. You make marker sized 3x3, then needed pixel is 9 pixel minimum probably.

If you perform detection using any kind of machine learning, then it depend to machine learning you use. The best way to get answer of your question is by testing.

$\endgroup$
1
$\begingroup$

As @jsotola has implied in the comments, context is everything. I think there will be many answers to your question in its general form. You may need to expand on your question to generate more meaningful answers.

Resources I found that might help you progress - -

Information about image processing in robotics. https://link.springer.com/chapter/10.1007/978-3-319-62533-1_12

Research from fields where they are forced to use limited resolution. “Image recognition with a limited number of pixels for visual prostheses design” (Sheng Li et al. Artif Organs. 2012 Mar) https://pubmed.ncbi.nlm.nih.gov/21954832/

The question seems related to your previous question Number of pixels on an object. Is that correct? Do they share context?

$\endgroup$
1
  • $\begingroup$ Hi @RowanP, Thanks for you response and the resources. Yes, you are right this is related to my previous question. And what I am trying to do is detect objects and discern what they are i.e. is it a monkey or a sheep at a distance so I figure there is a minimum number of pixels which would allow me to do this like say 28*28. This would be done out in the wild so the scene would not be simple. $\endgroup$
    – qwertyuiop
    Commented Jul 26, 2021 at 9:32

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.