First try with instance segmentation: Issue with number of instances in prediction

[jeanpat]

Hello,
I finally to perform to run instance segmentation on
a personal dataset (overlapping chromosome) in coco format.

The dataset consists in pairs of overlapping chromosomes, that is in each image, there are two instances of the same class of object (chromosome).

The model has two classes:

print(datamodule.labels)
print(datamodule.num_classes)
#help(datamodule.viz)
['background', 'chromosome']
2

When a prediction is performed, I get only one mask, for example:

I was waiting two masks. Does my dataset need to have three classes with background, chromosome1, chromosome2 ?

Thanks for your help

[ethanwharris]

Hi @jeanpat thanks for reporting this.

Does my dataset need to have three classes with background, chromosome1, chromosome2 ?

I think so but not totally sure. Perhaps @Borda may have some insight from working with instance segmentation for this notebook: https://www.kaggle.com/code/jirkaborovec/cell-instance-segm-lightning-flash