Incorrect implementation of "v" in "compute_ciou" function

[sanghyun518]

Describe the bug

In the current implementation of v in compute_ciou function (keras_hub/src/bounding_box/iou.py), the (4 / math.pi**2) term is included inside power and so gets squared together with the arctan term.

v = ops.squeeze(
ops.power(
(4 / math.pi**2)
* (ops.arctan(width_2 / height_2) - ops.arctan(width_1 / height_1)),
2,
),
axis=-1,
)

However, according to equation (8) in the original CIoU paper, (4 / math.pi**2) should NOT be squared (i.e., it should be outside the power op) - only the arctan term should be squared. This is correctly reflected in Ultralytics' implementation of CIoU here.

[harshaljanjani]

Hello @sanghyun518, thank you for pointing out this issue! You're right, the code does in fact show inconsistencies with Equation 8 in the research publication.
I just wanted to check with you and a collaborator if I could open a PR to fix this bug, unless you were already planning to do so.
This would be my first time contributing to keras-hub, although I am a contributor to the main keras repository.

[sanghyun518]

Hi @harshaljanjani , thanks for the kind offer but I was planning on creating a PR myself.

[sonali-kumari1]

Hi @sanghyun518 -

Another PR has been raised for this issue.

[sanghyun518]

Hi @sonali-kumari1 - I think this issue (in keras-hub) should be closed as the fix should be applied in keras-team/keras .