tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.27k stars 630 forks source link

some ideal about retinanet #295

Open buaacarzp opened 4 years ago

buaacarzp commented 4 years ago

Hello, the paper has discussed the method of centernes for Retinanet. It is mentioned in the paper that: "Centerness cannot be directly used in RetinaNet with Multiple Anchor Boxes per location because one location on Feature maps has only one center-ness score but different Anchor boxes on the same location require different "centerness" (note that center-ness is also used as "soft") For positive/negative samples). For anchor-based RetinaNet, the IoU score between anchor boxes and ground-truth boxes may serve as an alternative of "center-ness". I would like to ask you if this IOUscore cannot be calculated using L, T, B and R? Since the centers of all 9 anchors are the same, why not take all 9 anchors as positive samples?I don't quite understand that,please tell me . thanks!