Compared with the traditional anchor based method, fcos is unique

Thesis link ：FCOS original text
Paper code ：FCOS Code

FCOS yes anchor-freed one-stage Target detector , use first Backone CNN( Backbone architecture for feature extraction CNN), another s by feature map The total pace before

FCOS And anchor-based Different ways

• anchor-based The way is to use backbone,neck,head Obtained after calculation feature map, Then the regression classification is carried out through the preset a priori box , and FCOS obtain featuremap after , take featuremap The points on the map are mapped back to the original graph , The coordinate points mapped back to the original drawing are ：（floor(S/2) + x* s, floor(S/2) + y* s） floor Take the lower limit .
• anchor-based The way to judge a positive sample is ：anchor And GT Of IOU > threshold Set as positive sample , Category is this GT Category
  ,FCOS The way to judge the positive sample is to map back to the point of the original graph , If it falls on GT in , Is considered a positive sample , Category is this GT Category .
  Be careful ： The points mapped back to the original graph are easy to fall on multiple GT in , Such sample points are considered as fuzzy samples , You can use FPN Multi level prediction in , It will be mentioned later .
• Previous algorithms used multiple classifiers to classify , and FCOS use C A binary classifier to classify .（C Is the number of categories ）
• anchor-based Put different sizes of anchor Feature maps assigned to different levels . and FCOS the reason being that anchor-free Method , It directly defines different levels feature map Regression range , If the feature map is large, it will return to the small target object , On the contrary, if the feature map is small, it will return to the large target object , This solves the problem of fuzzy samples .
• FCOS In the proposed center-ness, Because after multi-level prediction, it is found that FCOS There is still a certain distance from the anchor frame based detector , The main reason is that there are many low-quality prediction frames far from the center of the target ,FCOS A formula of centrality is proposed , as follows ：
  
  According to this formula, we can get , Far from the center of the object center ness low , vice versa ,centerness Will be in loss When calculating * cls loss, Such a low-quality prediction frame will be in NMS It's filtered out when it comes to .