Odd performance stats

[asmagen]

I'm training a multi-class segmentation model based on adaptation of the tutorial from segmentation_models and the performance stats seem odd in the sense that while the predictions on the test set appear to be pretty good. Performance and example predictions on the test set: (1) the iou score seems to be better on the validation set (annotated in the figure as test) relative to the train set, which is unexpected, and (2) the loss on the validation set seems to not decrease by much, while the loss on the train set is decreasing, suggesting potential overfitting (but the performance on the test set is not bad)

Any ideas for what might be happening?

[JordanMakesMaps]

Do you think it could be how you're splitting your data up? (1) implies underfitting, so are you applying a lot of augmentation?