Pretrained model not working well on my data - struggling with flake detection and glue residue issues

[GarfieldofMaxwell]

I really love this project—both its overall design concept and the demonstrated results are absolutely amazing.So I've been trying to set up an automated system using the pretrained MaskTerial model, but I'm running into some issues. The model isn't performing well on my data at all—most of the time, it barely detects any flakes, and when it does, it's also picking up a lot of glue residue that I need to exclude.
What's more, even for some high-quality photos where large graphene flakes are clearly present, the model still doesn't recognize them properly. It's pretty confusing because these are obvious, good-quality samples.
I was wondering if this might be because the lighting in my images is a bit uneven?
Do you think fine-tuning the model on my specific dataset would help fix these problems? Or are there other adjustments I should try first? Any advice would be really appreciated!

[Jaluus]

Hey!

Thanks for using MaskTerial!
Yes, it is best to finetune the model on a few example images, about 5 to 10 per thickness should be good!
If you need examples, check out the dataset I used in the paper, there I have a few good examples.

About the flakes and images:

Hmm, yes the flake is quite visible, but very faint, are you using ~290nm SiOx by any chance?
The contrast is typically very faint and much better for 90nm SiOx!

Also your image has a few smudges and there seems to be dust in the optical column, I'd take a flatfield image and then subtract that to get a clean image, this should also increase the models performance!
This will also even out the lighting.

Typically the lighting should not be the problem for the Instance prediction model but rather the model which uses the contrast to predict the class.

Also, the instance prediction model has a prior on the flake sizes, but your flake size should still be within the expected size.