the ciou loss vs. computeciou

[ginobilinie]

Thanks for sharing the code.

I have noticed that in your other repositories (for example, https://github.com/Zzh-tju/DIoU-pytorch-detectron), you implement the computeciou function as the ciou loss, which is different from the implementation in this repository. I'm wondering why the implementation is different and which one do you suggest to use?

[ginobilinie]

OR ciou is only for ssd/yolact and computeciou for faster-rcnn/mask-rcnn?

[Zzh-tju]

The difference is due to the different variables feed in the function. You just choose them according to your basic model.

[ginobilinie]

I see. Thanks.

[kelvinkoh0308]

Hi,
In compute_ciou function, you implemented as
with torch.no_grad():
S = 1 - iouk
alpha = v / (S + v)
But in ciou function, you implemented as
with torch.no_grad():
S = (iou>0.5).float()
alpha= S*v/(1-iou+v)
Can I know what is the difference between these two?

[kelvinkoh0308]

And also which one is suggested to follow

[Zzh-tju]

@kelvinkoh0308 both are ok

[kelvinkoh0308]

@Zzh-tju thanks for your reply

[kelvinkoh0308]

@Zzh-tju can I know what is the reason that you add S = (iou>0.5).float() ?