RetinaNet, SSD, YOLOv3, Faster R-CNN are all Anchor-based detectors, i.e. predefined Anchor boxes need for training. FCOS is an Anchor-free and Proposal-free detectors, i.e. predefined Anchor boxes do not need to be trained, thus saving computing resources occupation.
As I understand it, the difference between FCOS and Faster R-CNN mainly in:
(1) Faster R-CNN each point is presented Proposal of N, N typically equal to 9 (the default aspect ratio of three kinds, three sizes thus Proposal 9), how each category Proposal sure? You need to calculate the Proposal and GT Box of the IOU, IOU reach a certain level, even if the Proposal is the GT Box class, this Proposal requires a return to the GT Box; and FCOS each point just made a Proposal, Proposal of how this category determine? This is the point at which a GT Box inside, Proposal on which it proposed a GT Box class belongs, of course, this Proposal requires a return to the GT Box. But the feeling Faster R-CNN classification better training, after all, Proposal and GT Box has a relatively large IOU, FCOS point near the center of the GT Box class training is still relatively good, relatively far from the center point estimate is not good training it.
(2) FCOS more out of a Center-ness:
This is because the lower the farther away from the center of the target predicted Proposal quality, so FCOS added a branch prediction Center-ness, closer to the center of GT Box, this Center-ness, the more the value of close to 1.
When the predicted binding and the classification score corresponding to the value of Center-ness to sort Proposal, thereby suppressing farther from the center position of the predicted target Proposal, this method is very clever, worth learning.
