I am trying to model a multi-class image segmentation problem. On doing some literature survey,
f-1 score and Jaccard index are coming to be a preferable way of measuring accuracy.
My classes are highly unbalanced. Which metric will give the most accurate representation?
Also, my model is predicting poorly on most of the classes, so most of the times, the outcome for a certain class is 0, so I can’t use Jaccard index in those cases as they’ll give undefined values according to this https://scikit-learn.org/stable/modules/generated/sklearn.metrics.jaccard_score.html
What should I do?
Also, any suggestions on how to improve the model performance, it is really bad and I don’t have any clue , how to improve it?