Score calculation as a product of the bbox and maskiou probability

[pieterblok]

@dbolya Hi Daniel, I found something strange in the postprocess function (in layers/output_utils.py). I'm using the yolact_plus_base_config settings.

In the first section of the function there is a filtering on the score_threshold (line 43). Later in the function there is a recalculation of the scores using the maskiou_p (line 86).

Shouldn't it be better that the filtering on the score_threshold is done after this calculation (applying the filer on the product of the bbox and maskiou probability)?

[dbolya]

Good catch, but that threshold was only meant to be a preliminary one to reduce processing time. The real threshold happens in prep_display, which in the YOLACT++ case would threshold again using the reweighted score. Since it's not possible that the score increases from rescoring, this shouldn't be an issue.

Though, there is a question of how all-in-one the postprocess function should be. I intended score_threshold to be there just for extra speed and for the user to threshold the score themselves, but looking back that makes no sense huh. I guess then we should fix this after all?

Also, I think your fix is missing the application of keep to the masks, since they get generated before the score reweighting (since they're necessary for that). You probably can just change

masks   = dets['mask']

on L104 to

masks   = masks[keep]

but I'd need to test.
(alternatively add the masks back to the dets dict like you did with the scores).