0

I have implemented a few algorithms for multi-class semantic segmentation and am now at the stage where I will be comparing them against each other and evaluating the results. I was wondering if there are any best practice methods/formulae for comparing segmentation, and their advantages etc?

So far I've looked into the problem of class imbalance pointing out the need for something more complex than just pixel counting, which has led me to the Sorensen-Dice coefficient - while this seems to be appropriate for single class scenarios (which I could apply to my current problem), I am looking for something more directly appropriate for the task at hand. Thanks

Note: The method does not need to be fast/efficient or run in real time, it is just the results I am after.

Gabriel Lacey
  • 3
  • 3
  • did you work out any solution for the multi-class dice coeff? – AI_NA Apr 28 '20 at 19:38
  • @AI_NA Yeah I ended up doing that for the final project. I essentially get the dice coeff for each region individually in the image (by treating all other regions as background), and once I have a score for all of them I average them all out to get a dice coeff for the image as a whole. Not ideal, but it does allow me to compare different images. – Gabriel Lacey Apr 29 '20 at 20:51

0 Answers0