0

I have been looking around into papers like Pix2Pix and DeeplabV3. And In pix2pix paper there is an image to segmentation task performed. similarly, DeeplabV3 is used for segmentation tasks. I want to understand when should I use GAN for segmentation and when should I use SOTA segmentation methods. What are the pros and cons of these methods on the semantic segmentation?

techno
  • 101
  • 1
  • 1
  • 8

1 Answers1

1

I guess someone may find an example but I haven't seen a GAN method performing better than the state of the art segmentation methods so far. In my understanding GANs are more interesting for domain transfer, or weakly supervised methods in general. However, if you have a set of data with corresponding pixel annotations, I think that you should follow the "standard" semantic segmentation papers.
If you want to know which methods perform better in general, I would recommend you to follow some computer vision challenges and check for the best performing methods (e.g. kaggle, grand-challenge.org, conference challenges, ...).
One website is interesting to keep track of different methods on a single graph: https://paperswithcode.com/task/semantic-segmentation
Therefore if you are wondering what would work best for semantic segmentation, I would certainly advise you to look at deeplab as you mentionned or this repository, quite effective.

87VN0
  • 775
  • 3
  • 10