Ciou loss

In the upper right corner is 3 In this case , Green is gt box, Black is the prediction box , You can see three situations l2 The loss is the same , but iou It's a lot worse , So lead to iou loss,iou loss=-ln(iou).

giou loss

You can see , The green in the lower left corner is gt box, Red is the prediction box , Blue is the formula in the upper left corner Ac 了 , In the formula u It refers to the union of green and red ,
giou loss=1-giou,

diou

The first line of the left figure is using giou As loss value （ Black is grid cell, Green is gt box, Blue is the prediction box ,400 Times barely coincide ）, The second line is to use diou As loss value ,120 Times have coincided perfectly .


In the formula b It refers to the central coordinate of the prediction box ,bgt refer to gt box The coordinates of the center point of the ,

ciou

The aspect ratio is introduced .

focal loss

Generally, the detection algorithm in one stage will encounter the problem that the number of positive and negative samples does not match . You can see that the red boxes in the upper right corner are all negative samples , A lot .