apply dIOU replace gIOU

[lucasjinreal]

Will dIOU obtain a better result then gIOU?

[glenn-jocher]

I don't know. You can use g/d/ciou in the loss function by specifying it here. To use DIoU you would pass DIoU=True in place of the current GIoU on L413.

https://github.com/ultralytics/yolov5/blob/111994956ede6964433e02bc3f64deb04b9916ac/utils/utils.py#L409-L413

[aktiver]

Update for Yolov5: This is now inloss.pyon line 141:

Default (CIoU): Paper iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze() # iou(prediction, target)

GIoU: Paper iou = bbox_iou(pbox, tbox[i], GIoU=True).squeeze() # iou(prediction, target)

DIoU: Paper iou = bbox_iou(pbox, tbox[i], DIoU=True).squeeze() # iou(prediction, target)

The definition of bbox_iou is on line 216 in metrics.py.

[glenn-jocher]

@aktiver thanks for your update! The implementation of DIoU, GIoU, and CIoU are all available in the loss.py file, and the bbox_iou definition can be found in the metrics.py file. You can now easily choose and use the desired IoU metric in your YOLOv5 implementation. If you have further questions or need assistance, feel free to reach out.