Directly run yolo score will not match your paper.

[guyuchao]

I directly run the following command "yolo val segment model=yolov8m-seg.pt data=coco.yaml device=0".
Note I don't use instance diffusion output, but use original coco data. Why it doesn't match the performance of Upper bound (real image) in your paper: