Why regression head wasn't normalized ?

[RocketFlash]

Thank you for the amazing work! In the last FCOS model update, you divided regression head predictions by the FPN level stride value. However, why not divide regression predictions by the image size instead? In this case, all values would still be positive, and it would be possible to use a sigmoid activation at the regression output. I assume you did this intentionally, so I would like to understand the reasoning behind it.