Intersection over Union (IOU), Generalised IoU ( GIOU ), Distance IOU (DIOU) & Complete IoU (CIOU)

Hi everyone. This video discusses the various IoU metrics being used for the task of Object Detection, namely , IoU, GIoU, DIoU and CIoU.

Please like and subscribe to this channel if you like the content and share with your friends.

---Contents----
0:00 - IoU
2:06 - GIoU
6:02 - DIoU
7:53 - CIoU

[abhishektodkar]

This video should reach out to more people..

[mohammadyahya]

You mentioned at 4:20 that the only only to minimize GIOU loss when there is no overlapping between predicted and ground truth bounding boxes is by increasing the bounding box area so that it becomes bigger and thus occupy more red space. But why not the only way to minimize the red distance is by making the predicted bounding box closer and ground truth closer together instead of making predicted bounding box bigger please?

[mohammadyahya]

Why negative gradients in GIoU in 3:10 would make the predicted bounding box moves toward the ground truth bounding box please?

[qjhlxbq]

Just to clarify for the GIoU concept, we want
- a higher IoU between the ground truth bounding box and the predicted bounding box
- a lower/smaller area in the C box (excluding A union B)
- a smaller C box (excluding A union B) divided by C
- a higher GIoU => from a high IoU and low 'smaller C box (excluding A union B) divided by C'
- a lower Loss_GIoU
?

[mohammadyahya]

For DIOU, what if the predicted bounding box and the ground truth bounding box centers are the same at 6:49? In this case, we will have similar case to GIoU where Loss IOU = Loss DIoU and thus we will have no gradients flowing back, is that correct?

[mohammadyahya]

At 7:40, you mentioned that for the horizantal/vertical case, the distance will be larger than area thus DIoU is better that GIoU although I can see that the distance is not very large, please correct me if I am wrong?

[mohammadyahya]

Why for 5:56 in the case of innacurate regression, there will be no gradient to flow when Loss IoU and loss GIoU are the same?

[mohammadyahya]

Sorry for asking so many questions, but I am new to this area. At 4:30, you mentioned predicted bounding box in blue, anchor box in black and green box is the ground truth. What is the difference between predicted bounding box in blue and anchor box in black please?

[mohammadyahya]

Thank you again. To overcome gradients not flowing in IoU, you mentioned that GIoU solves this. In case two boxes does not overlap, we will have CIoU = IoU - rest. IoU=0 due to not overlapping, so we will have CIoU to be minus number (minus gradient). How we can update the weights with negative gradients please? Is this possible?

[mohammadyahya]

Last question pleased, how CIoU will take care of overlapping issue when the predicted bounding box is smaller but at the same center as the ground truth bounding box?

[mohammadyahya]

May you please clarify more why IoU is scale invariant?

[mohammadyahya]

For GIoU, in case A takes half of C and B occupies the other half where A and B has 0 intersection, then Loss GIoU = 1 - (0 + 0) = 1. So since this is not 0 but 1, this is how GIoU avoids no gradients flowing back if there is no intersection please?