I am somewhat confused on the terminology used in Computer Vision. Suppose I wanted to detect (localize) an object (say, a human) in an image. However, a box around the human is not sufficient, what I need is a pixelwise localization. More precisely, I need a mask telling me which pixels belong to the human and which ones don't. If I wanted to find papers dealing with this problem, what would be the term to search for?

  • Object detection: Yes, but most approaches actually work in terms of bounding boxes
  • Object localization: If I see it correctly, this is essentially used as a synonym for object detection
  • Image segmentation: Here we segment objects in terms of pixels but we do not necessarily only aim to segment out one specific object that is known.

Is there very famous work in this area I should not miss?

  • 1
    $\begingroup$ "what I need is a pixelwise localization" For what? $\endgroup$
    – endolith
    Feb 19, 2016 at 16:55
  • 1
    $\begingroup$ How is a box not "pixelwise"? A location is a single point. $\endgroup$ Feb 20, 2016 at 15:58
  • 1
    $\begingroup$ Define very precisely what you actually need; that might be hard, maybe start with an actual example of an image as input, and an actual representation of the output you want. $\endgroup$ Feb 20, 2016 at 16:01
  • 1
    $\begingroup$ You need to look for "Human Pose Estimation" and of course Segmentation (if all you need is the pixels belonging to the human). $\endgroup$ Feb 20, 2016 at 20:42
  • 1
    $\begingroup$ @user1809923: obviously, it is not obvious what you want. That's the whole point of this comment discussion... $\endgroup$ Feb 22, 2016 at 11:07


Your Answer

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