I have source data that can be represented as a 2D image of many similar curves. They may oftentimes cross over one another, so regions of interest will overlap.

My goal is to implement a neural network solution to identify each instance or the curves and the pixels that are associated with each instance.

(Each image is simple in its representation of the data. A pixel in the image is either a point on one of these curves or it is empty. So the image is represented by one or zero at each pixel. For training purposes, I have labels for every pixel, and I have about 150,000 images. The information in the images can be noisy in that there may be omissions of points and point locations are quantized due to measurement limitations and preprocessing for the image preparation.)

I started looking into what semantic segmentation can do for me, but since all of the instances are of the same class, distinguished mainly by their location in the image, I don't think semantic segmentation is the type of processing I would want to perform. (Am I wrong?)

I am very interested in seeing how a neural network will work on this problem to separate each instance.

My question is this: what is the terminology that describes the process I'm looking for? (How can I effectively research for this problem?) Is this an extension of semantic segmentation or is it referred to some other way?


What you want to do is instance segmentation on a pixel level. I can point you two different way:

  • $\begingroup$ Thanks. This looks like exactly the path I want to follow in my investigation."Instance segmentation" looks to be the right terminology to begin searching. The papers look very helpful. $\endgroup$
    – Jim
    Jan 17 '19 at 19:41
  • $\begingroup$ Also look for YOLOv3 which is a single shot detector. $\endgroup$ Jan 19 '19 at 11:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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