Is GIOU or DIOU or CIOU better than IOU?

[ZhengMengbin]

As mentioned in title, have you used giou/diou/ciou to train yolox, and who is better?

[debapriyamaji]

@ZhengMengbin , I think, this repo has support for iou and giou loss. By default, iou loss is used. https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/models/losses.py

BTW, where is it mentioned that giou/diou/ciou loss is used?

[MengtianLee]

I have used giou, ciou and eiou to train yolox-tiny. Unfortunately, no one is better than default iou loss and i found that the learning rate of them is significantly small. BTW, I will try increase the initial learning rate of them.

[GF1447571253]

How it work

[doctorchoco]

what was the result of the experiment?

[MengtianLee]

It doesn't work by increasing the learning rate.

[CarlHuangNuc]

It doesn't work by increasing the learning rate.

so the conclusion is CIOU /DIOU is not better than IOU loss in YOLOX ?

[wgq18]

After I use giou or siou, I get nan in iou loss and give me error 'Assertion input_val >= zero && input_val <= one failed.' How did you fix this error? @MengtianLee

[twmht]

Any update?