Tensorflow implementation for Distance-IoU Loss from the Paper
The function return the ciou of two lists of bounding boxes, so when using apply tf.reduce_mean(), or whatever way to optimize over the ciou returned.
on using the ciou or diou loss, I found that -log(IoU) is more stable and converge faster than (1-IoU), so in that case ciou will become -log(IoU) + u + alpha*ar , also I would suggest using ciou along with bounding box regression using Huber loss.
