We're having trouble making working layer separation models for MS73 (possible reasons listed)

[JoyfulGen]

At the moment, @kyrieb-ekat and I don't have a working model for MS73. Models that we make work only for the folios used to make them; if we try to separate layers on a folio the models were not trained on, it doesn't work; the layers are all mixed up, or a layer is almost empty, or a layer is super speckly.

Here's an example of a model trained with folios 054 and 288 separating layers on folio 274:

We don't know exactly why this is happening yet, so I have gathered theories/questions. Each theory/question has a separate issue detailing it and giving examples. Issues 2 and 3 aren't confirmed yet, but I've put them up anyway to gather data. @kyrieb-ekat, please add any other relevant thing! Or remove irrelevant things!

• Background_removal is removing the middle of neumes and text: Background removal isn't working correctly #1201
• Our models might be overfitting (new word learned by me): MS73 layer models might be overfitting #1202
• Using half of a folio image in Pixel produces different results than using the full image and selecting half: Training with half an image produces different results than with a selection of that same image #1203

[kyrieb-ekat]

To build on this- models will do very well for a couple rounds and then stop working and perform poorly. For example models trained up very nicely on multiple folios will do well on a chunk of the manuscript, and then after correcting and retraining stop producing layers, or start separating the background incorrectly.

[kyrieb-ekat]

Deleted additional comments to open a new issue #1206