How should the data for training look like?

[Tentenn]

I have hand annotated segmentation data. (image - mask pairs in png format) and want to train the model to make it better

I see that the training scripts makes use of 4 folders: fg/, bg/, masks/, imgs/

What images go into each folder?

• bg/: That's clear, only background images
• fg/: Should there be images of only characters and white background or transparent?
• masks/: Is this the silhouette of the images in imgs/? Do the names have to match? Is there a naming convention?
• imgs/: No idea what goes in there. Maybe raw images with foreground and background?

Are there any constraints to the data? (pixel size, image size)

Thank you for your work!

[SkyTNT]

You're almost right.

Did you check out my dataset https://huggingface.co/datasets/skytnt/anime-segmentation ?

[Tentenn]

Thanks! I checked it out now. It makes more sense now.