What is the difference between 'In GT Box' and 'Spatial constraint' about atss HOT 4 CLOSED

@leeesangwon

1. As for RetinaNet + imprs. (Intersection over Union) and FCOS (Spatial and Scale Constraint) in table 2, their final positive samples are limited in GT box. But the spatial constraint is a step for FCOS to select candicate positives. The 0.8 performance gap comes from how to select final positives from those anchors in GT box across all pyramid levels.
2. If allow_low_quality_matches=False, you are right, but allow_low_quality_matches=True.

from atss.

@leeesangwon
Spatial Constraint can ensure In GT Box, while IoU can not, i.e., IoU may select some positive anchors whose center are on the outside of GT box, thus we limit positives in GT box.

from atss.

1. Then what I understood are:

   • Both RetinaNet + imprs. (Intersection over Union) and FCOS (Spatial and Scale Constraint) in table 2 have the spatial constraint.
   • The 0.8 performance gap comes from whether using all candidate boxes on the corresponding pyramid level or selecting boxes using the IoU threshold.

   Did I understand correctly?

2. If I did, I have one more question.

   As you commented, the IoU threshold can not ensure the matched anchor box is in the GT box.
   However, I think it can if the IoU threshold is large enough.
   For example, RetinaNet uses 0.5 as the threshold and it seems sufficient condition. (Please refer to my low-quality figure below)
   

   But you could improve performance after limiting positives in the GT box.
   Please let me know where is the points I missed!

Thank you!

from atss.

Now everything becomes crystal clear!
Thanks again your super fast reply!! 🚀

from atss.