What is CIOU loss?
CIOU introduces two new concepts compared to IOU loss.
The concept of centroid distance, which is the distance between the centroid of the actual bounding box and the centroid of the predicted bounding box.
The concept of comparing the aspect ratio of the actual bounding box with the aspect ratio of the predicted bounding box.
Through those 3 measures, we can evaluate the quality of the predicted bounding box.
So, let’s see the formula for CIOU loss.
Formula of CIOU
$Loss_{CIoU}\ =\ 1\ -\ IoU\ +\ \frac{p^{2}(b,b^{gt})}{c^{2}}\ +\ \alpha v$
$Loss_{CIoU}\ =\ 1\ -\ \frac{Area\ of\ Overlap}{Area\ of\ Union}\ +\ \frac{p^{2}(b,b^{gt})}{c^{2}}\ +\ \alpha v $
where:
$B^{gt} = (x^{gt}, y^{gt}, w^{gt}, h^{gt})$ is the ground-truth
$B = (x, y, w, h)$ is the predicted box.
$c$ = the diagonal length of the smallest bounding box.
$p$ = Euclidean distance.
$v = \frac{4}{\pi}(arctahn\frac{w^{gt}}{h^{gt}}\ -\ arctahn\frac{w}{h})^{2}$ = measures the consistency of the aspect ratio.
$\alpha= \frac{v}{(1\ -\ IoU) + v} = \frac{v}{(1\ -\ \frac{Area\ of\ Overlap}{Area\ of\ Union}) + v}$