How is accuracy defined for semantic segmentation?

How is accuracy defined for semantic segmentation?

@Meir
Sematic Segmentation can been seen as multi-class classification problem, right? You are trying to identify a particular pixel belongs to a class A vs class B, etc… especially if you think the classes are represented in one-hot encoding. It will be True if predict the correct class, or otherwise False.
That said, you can look at this as a multiple binary classification problem therefore for the same logic, you can use binary cross entropy loss for each class and sum over all the classes. (or you can also take the mean.)

Does it make sense? Let me know.

Suki

So, you basically view each pixel as a separate test instance? And the accuracy is the total number of pixels for which the correct class is predicted over the overall total number of pixels in all images in the test set?