How to understand "reduce_zero_label" in MMSeg

[ChubbyPear]

I have a custom dataset with a background and 4 classes. In my annotations, 0 represents the background, and 1 to 4 represent the classes. The data distribution is highly imbalanced, with over 90% being background.

When I set reduce_zero_label=False and num_classes=5, the model identifies the entire image as background. However, when I set reduce_zero_label=True and num_classes=4, the model performs better but misclassifies the background as another class.

I’m confused about how the models in MMSeg handle the background when reduce_zero_label=True, especially since "background" isn't defined as a class. During inference, my model incorrectly classifies the background as one of the specific classes (one from class 1 to 4). How can I address this issue to ensure the background is correctly identified?