Perform multi-class segmentation with pixel-labels, which only contain single class

For example, we have class number 3 in total.
But the training set only contains classes like, [1,0,0],[0,1,0], and [0,0,1]
What can i do to improve the performance?

The number of targets matches the number of classes, so I’m unsure what the issue is.
In case you are struggling to get any improvements in your model training I would recommend to try to overfit a small dataset (e.g. just 10 samples) by playing around with some hyperparameters first.

But there is no traning samples containing multi-class.
When inference, there may be [1,1,1] containing 3 class, it not work well.
I try to use cutmix to increase the diversity of training samples, but it improved a little.