Adding DIoU loss and CIoU Loss to keras-cv

[Anirban0011]

Should this two losses be added to keras_cv.losses ?

1. DIoU (Distance over IoU) Loss is independent to the size of the object being detected. It is seen to converge mush faster than GIoU even with opposite orientation of boxes(vertical and horizontal). It is also deployed as a criterion in NMS layer of which we already have an implementation in the repo.

   • paper link : https://arxiv.org/abs/1911.08287
   • implementation link : https://github.com/Zzh-tju/DIoU

2. CIoU (Complete IoU) loss needs fewer iteration than both GIoU and DIoU loss to converge. It has been deployed in YOLO v3, Yolo v4, SSD, and Faster RCNN.

   • paper link (same) : https://arxiv.org/abs/1911.08287
   • implementation link : https://github.com/Zzh-tju/CIoU

[Anirban0011]

A paper which highlights the limitations of existing BBR loss functions : Focal and Efficient IOU Loss for Accurate Bounding Box Regression

[jaygala223]

Hello @ianstenbit, can you please take a look at this issue? I would like to work on this