When I label images for semantic segmentation (using u-net, if that matters), is labeling the background (anything I am not interested in) necessary? Will it improve the network's performance?

  • 1
    $\begingroup$ I am not sure I understand your concern. In image segmentation, you want to label each pixel. So, you will have the labels for each of your objects of interest. I think the remaining objects will just be considered background. If you explicitly label the background, then I suspect that this could mean that you force the background to look in a certain way, but I am not sure about the implications. To be honest, I am not sure what the benchmark datasets like COCO do. You may want to investigate that. $\endgroup$ – nbro Jul 9 '20 at 17:09
  • $\begingroup$ Maybe this dataset ADE20K is worth investigating too. $\endgroup$ – nbro Jul 9 '20 at 17:17
  • $\begingroup$ Thanks, that's interesting. I'll see if I can find the time to benchmark that $\endgroup$ – liorr Jul 9 '20 at 23:15

Your Answer

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

Browse other questions tagged or ask your own question.