How to understand the output of different tasks

[Zhong-Zi-Zeng]

Hello, I am a beginner in instance segmentation and panoptic segmentation. I want to figure out the output of MaskFormer.

In semantic segmentation tasks, classes with the same predicted probabilities will have their corresponding masks merged as the final output.

In instance segmentation tasks, it is sufficient to filter out the masks corresponding to probabilities below a certain threshold.

In the case of panoptic segmentation, it is necessary to consider the corresponding data format. Taking COCO panoptic as an example, it contains 80 "thing" classes and 53 "stuff" classes. So, should the number of queries be set to 133, or is there another method to determine it?

I would greatly appreciate it if you could help me clarify these questions. Thank you.