I found a bug that severely affects your model

[18112330636]

You are using the giou of the outer frame during training, but this is obviously a problem. If the center of a square is rotated ten degrees to the left and ten degrees to the right, the poly IOU difference is very large, and the IOU difference of the outer frame is very small, which will affect For the convergence of the model, it is recommended to change the IOU of the outer frame to poly IOU or rotate IOU

[18112330636]

Can you put the angle directly into the IOU and use the latest KF IOU

[18112330636]

Moreover, it is wrong to use the IOU of the outer frame when calculating the map, and the IOU of the two identical frames with a certain angle to the center left and the center right is problematic.

[hukaixuan19970627]

In validation yolov5_obb caculate the HBB mAP for faster training, you can get the obb mAP by using DOTA_devkit/dota_evaluation_task1.py. For example:

[hukaixuan19970627]

And I think you may have some misconceptions about this repo. In yolov5_obb：

    CIoU Loss -> (cx, cy, w, h)
    Circular Smooth Label+BCELoss -> (θ)
    BCELoss(pred, detached CIoU ×gt) -> (obj)

Yours:

    R-IoU/R-GIoU/.../KFIoU Loss -> (cx, cy, w, h, θ)
    BCELoss(pred, detached R-IoU ×gt) -> (obj)

Both can be used to detect oriented objects.