Currently my understanding of dice score (in that task of semantic segmentation) is that
we iterate over our input images
for each class
we calculate the dice score the ground truth mask of that class and the predicted mask
now my question is what happen when the ground truth mask and the predicted mask of certain class are both empty, do we skip the calculation for the this class or do we reward the model for not predicting any false positive by setting the score to 1 directly