Understanding of IOU GIOU DIOU CIOU in target detection

IOU paper: link
GIOU paper: link
DIOU paper: link
CIOU paper: link

The original IOU has the following problems:

1. The general two-stage network frame regression IOU≥0.5, the frame will not be returned
2. If there is no overlap, it is always 0 and cannot be optimized.
3. Can’t reflect how the two boxes intersect

GIOU can solve problems such as 1, 2 and can give a gradient value. A penalty term is added on the basis of IOU. When the distance of the bbox is larger, the penalty term will be larger.
GIOU problems:
1. The two boxes contain At that time, GIOU will degenerate into IOU
2. GIOU needs to iterate many times to converge

DIOU proposes the normalized distance of the center point, which can solve this kind of problem. The
paper proposes that DIOU is used in nms, namely DIOU-NMS , and the test results show that there is a certain improvement (improvement direction!!!)

CIOU adds a penalty factor to DIOU, alpha*v, this alpha is a weight function, v is used to measure the similarity of the aspect ratio

For faster rcnn:

1. The performance of IOU, GIOU and DIOU is very good!
2. The effect on medium and large objects: GIOU has very small gain, DIOU and CIOU work very well, CIOU is especially good (considering the length and width and the center point, the effect is good!)
3. Situation on small objects: CIOU is a little better than IOU, and DIOU is better. The possible reason is that the center point is more important than the length and width.
   Improvement direction: faster rcnn+DIOU faster rcnn+CIOU

References:
Target detection regression loss function: https://zhuanlan.zhihu.com/p/104236411
Sorting out various losses: https://blog.csdn.net/qq_14845119/article/details/80787753
GIOU: https:/ /mp.weixin.qq.com/s/CNVgrIkv8hVyLRhMuQ40EA
AAAI 2020 | DIoU and CIoU: The correct way to open IoU in target detection: https://bbs.cvmart.net/articles/1396