[FEATURE] aggregrated global evaluations

[neuronflow]

Is your feature request related to a problem? Please describe. sometimes, multiple patches are evaluated that come from a bigger entity, e.g., multiple microscopy slices from one brain. computing aggregated statistics on a patch level leads to different treatment of blobs depending on their patch. E.g. one patch might gave 5 instances another 100, so the instances from the less populated image volumes have much bigger impact on the aggregated statistics.

Describe the solution you'd like a flag to aggregate globally, "across" image volumes.

Describe alternatives you've considered for RQ it is quite trivial to compute this from the confusion matrix, but for SQ it is not so easy.

[Hendrik-code]

How would you differentiate between patches of image A and a new image B?

1. Can you not stitch the image together before inputting it into panoptica?
2. If 1) fails, I guess what we can do is you can give the evaluate() call a list of predictions and reference arrays, being different patches of the same subject. Would still be janky, but I don't see how a simple flag in the evaluation helps this issue.

[neuronflow]

1. Can you not stitch the image together before inputting it into panoptica?

no they might just be samples from one bigger entity e.g. cropped slices from one mouse brain.