Semantic segmentation & Train with more data

[ghost]

Your model is the best one for Anime segmentation task that I have ever seen.

There is a new crowd-sourced dataset for danbooru. Hope it will help further improve the performance.

It would be even better if you could add more classes for the segmentation.

[zymk9]

Thanks for the advice! I did plan to expand and improve the quality of the dataset.

Have thought about bringing in more classes when starting this project, but it soon turned out to be unmanageable to do all the annotations myself. Yet even for the current task quite some performance gain is needed.

[ghost]

I also saw a annotation project called findpose. Combined with DanbooRegion it may help truning a lot of resources into useful semantic maps

[mrj-taffy]

good job! and I wonder konw how many manual segmentation data used to train the SoloV2? does the AniSeg Dataset matters?

[zymk9]

The raw dataset contains only around 1k2 annotations, where 900 or so are from AniSeg.