Question on finetuning

[hpichlerbio]

Thanks for the great work!

I am really interested in finetuning this on my own data, but I am not quite sure, if it is possible.
Can I also finetune the model on instance segmentation data? Semantic segmentation is not possible in my case.
If not, do you have any recommendations on foundation models capable of doing instance segmentation without having to add an adapter to the model?

Thank you in advance!

[theodore-zhao]

Thanks for the feedback! The model currently output semantic segmentation only. I will add a postprocessing script on the output to turn it into instance segmentation. Then you can finetune on semantic mask and post process to instance outputs.

[hpichlerbio]

Thanks for the feedback! The model currently output semantic segmentation only. I will add a postprocessing script on the output to turn it into instance segmentation. Then you can finetune on semantic mask and post process to instance outputs.

Thanks, that would be amazing!

[evanrubel]

Out of curiosity, is the post-processing very involved? Or does the model provide some kind of raw output to identify different instances?

[hpichlerbio]

Based on the aswer I was given, I don´t think the model returns anything regarding single instances. The post-processing would therefore probably not detect overlapping instances. But I´m not 100% sure

[theodore-zhao]

The postprocessing is a simple off the shelf tool to partition the semantic output into non-overlapping instance sub regions. Yes. the model doesn't output instances related information.

[theodore-zhao]

A pathology instance inference example is just provided in inference_examples_RGB.ipynb. There is a simple code for post-processing the semantic output. Hope that works for you!

[hpichlerbio]

Thanks, I will look into that :)