Hi all teams of Pytorch Community.
I have a enquiry regarding FCOS (Fully Convolutional One-Stage Object Detection) code in PyTorch. I have checked FCOS Pytorch subcode in Github, the author uses “Anchorgenerator” like Retinanet and other anchor-based detectors. while the FCOS net is the Anchor-free detector according to this paper.
do anyone know that why the programmer uses Anchor-generator in this code?