Is there any way to train only semantic segmentation dataset like ade20k?

[ET-LONG]

I've tried to change it, it might be related to the network, I can't directly train only the semantic segmentation dataset, but about ade20k some example data I can't download, and I want to know how to train only the semantic segmentation dataset.
For example

Setting the data in the first image to the second will result in an error inside the model.