There is something I'd like you to tell me

[さいさい]

When creating a ground truth of semantic segmentation, how much does it affect depending on whether you paint or paint object boundaries?
In other words, is it better to put a boundary?
Also, please let me know if you know papers on this.