I am working on multi-class image segmentation and currently having challenges regarding my dataset. The labels (ground truth/target) are already one-hot encoded for the two class labels but the background are not given. Firstly, is the annotation or labeling of the background necessary for the performance of the model since it will be dropped during prediction or inference?
Secondly, due to the highly imbalance nature of the dataset, suggest approaches as read on the forum is either to use weighted sampler or weighted loss function. As my preferred choice of weighted dice loss function, how will I compute the weight for each class (suggested approaches are targeted at classification task not segmentation)?
Finally, what is the significance of a combine loss function i.e dice loss and cross-entropy loss function?
I appreciate your explanations and suggestions in advance.