How does Mask2former do the box evaluation in instance segmentation?

[Agito555]

Question 1:
I notice that there is one line code evaluation = dict(metric=['bbox', 'segm']) in the config file mask2former_r50_lsj_8x2_50e_coco.py.
However, according to the paper, Mask2former doesn't need to predict the boxes to predict the mask. Actually, they don't predict any box, right?
I wonder what does the box evaluation mean in here?

Question 2:
Similarly, when I visualize the instance segmentation results by tools/test.py, there are lots of boxes in images. Are they used to do the box evaluation? Or are they related with something about the receptive filed of the query features?

@chhluo Any help is appreciated.

[Agito555]

I get the idea. The network doesn't need to output the box predictions. The box for evaluation is the minimal rectangle which covers the mask of object.