IOU Loss
I O U = a r e a ( p r e d ) ∩ a r e a ( g t ) a r e a ( p r e d ) ∪ a r e a ( g t ) IOU = \frac{area(pred) \cap area(gt)}{area(pred) \cup area(gt)}I O U =a re a (p re d )∪a re a (g t )a re a (p re d )∩a re a (g t )
缺点: 不相交时IOU loss为0.
I O U L o s s = − l n p r e d ∩ g t p r e d ∪ g t IOU Loss = -ln\frac{pred ~\cap~ gt}{pred ~\cup~ gt}I O UL oss =−l n p re d ∪g t p re d ∩g t
GIOU Loss
G I O U = I O U − A r e a − p r e d ∪ g t A r e a GIOU= IOU – \frac{Area – pred ~\cup~ gt}{Area}G I O U =I O U −A re a A re a −p re d ∪g t
− 1 ≤ -1 \leq −1 ≤ GIOU Loss ≤ \leq ≤ 1
其中 Area 是能够覆盖住 pred 和 gt 框的最小矩形的面积。
G I O U L o s s = 1 − G I O U GIOU Loss = 1-GIOU G I O UL oss =1 −G I O U
DIOU Loss
G I O U = I O U − d i s t a n c e ( p r e d c e n t e r , g t c e n t e r ) 2 c 2 GIOU= IOU – \frac{distance(pred_{center}, gt_{center})^2}{c^2}G I O U =I O U −c 2 d i s t an ce (p re d ce n t er ,g t ce n t er )2
− 1 ≤ -1 \leq −1 ≤ DIOU ≤ \leq ≤ 1
其中 c 是能够覆盖住 pred 和 gt 框的最小矩形的对角线。
G I O U L o s s = 1 − G I O U GIOU Loss = 1-GIOU G I O UL oss =1 −G I O U
CIOU Loss
三项分别考虑重叠面积，中心点距离，宽高比
C I O U = I O U − ( d i s t a n c e ( p r e d c e n t e r , g t c e n t e r ) 2 c 2 + α v ) CIOU = IOU – (\frac{distance(pred_{center}, gt_{center})^2}{c^2}+\alpha v)C I O U =I O U −(c 2 d i s t an ce (p re d ce n t er ,g t ce n t er )2 +αv )
v = 4 π 2 ( a r c t a n ( w g t h g t ) − a r c t a n ( w h ) ) v = \frac{4}{\pi^2}(arctan(\frac{w^{gt}}{h^{gt}})-arctan(\frac{w}{h}))v =π2 4 (a rc t an (h g t w g t )−a rc t an (h w ))
α = v 1 − I O U + v \alpha=\frac{v}{1-IOU+v}α=1 −I O U +v v
C I O U L o s s = 1 − C I O U CIOU Loss = 1-CIOU C I O UL oss =1 −C I O U
Focal Loss
F L ( p t ) = − ( 1 − p t ) γ l o g ( p t ) FL(p_t) = -(1-p_t)^\gamma log(p_t)F L (p t )=−(1 −p t )γl o g (p t )
p t = { p i f y = 1 1 − p o t h e r w i s e p_t = \left{ \begin{array}{ll} p & if ~y = 1 \ 1-p & otherwise \end{array}\right.p t ={p 1 −p i f y =1 o t h er w i se
( 1 − p t ) γ (1-p_t)^\gamma (1 −p t )γ可以降低易分样本的损失贡献。
3.1 YOLO系列理论合集(YOLOv1~v3)_yolov3 spp _Focal Loss讲解 33:10时刻
有时候实际代码实现中会乘多一个α t \alpha_t αt 来平衡一下loss的权重(( 1 − p t ) γ (1-p_t)^\gamma (1 −p t )γ)。作为一个超参数。
F L ( p t ) = − α t ( 1 − p t ) γ l o g ( p t ) FL(p_t) = -\alpha_t(1-p_t)^\gamma log(p_t)F L (p t )=−αt (1 −p t )γl o g (p t )
注意，数据集标注要比较准确，不然这些噪声会干扰focal loss下的训练。
并且使用时需要调好参。
Original: https://blog.csdn.net/weixin_43850253/article/details/126326530
Author: Andy Dennis
Title: 目标检测损失函数
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/681101/
转载文章受原作者版权保护。转载请注明原作者出处！