Missing function `cal_my_giou`

[rwang15]

When calling bbox_overlap_loss, it tries to load Rotated_IoU.oriented_iou_loss.cal_my_giou but failed and always fallback to _cpu_fallback_giou. It seems the function is not implemented in this repo or the original `` repo. Is that function forgotten? or just deprectated for some reason?

[Cylrx]

Hi, I've tried to implemented a work around for this issue by using the cal_giou method from oriented_iou_loss and adding the necessary init.py to resolve the import error. This change successfully reproduces the "bedroom-0" case from the paper (Figure 7, Column 1). However, results for many other cases remain unsatisfactory.

Also, I've noticed that the _cpu_fallback_giou method seems to be detached from the computational graph, which is weird.

(Image 1 is the result from the original implementation, Image 2 is the worked around version)

[YukiiSetsuna]

Hi, I've tried to implemented a work around for this issue by using the cal_giou method from oriented_iou_loss and adding the necessary init.py to resolve the import error. This change successfully reproduces the "bedroom-0" case from the paper (Figure 7, Column 1). However, results for many other cases remain unsatisfactory.

Also, I've noticed that the _cpu_fallback_giou method seems to be detached from the computational graph, which is weird.

(Image 1 is the result from the original implementation, Image 2 is the worked around version)

Hi, I’ve been struggling with the same cal_my_giou problem. Would you mind sharing a bit more detail about how you modified the code (e.g., how you used the cal_giou method )?

Any pointers or snippets would be really helpful.

[Cylrx]

Hi, I've tried to implemented a work around for this issue by using the cal_giou method from oriented_iou_loss and adding the necessary init.py to resolve the import error. This change successfully reproduces the "bedroom-0" case from the paper (Figure 7, Column 1). However, results for many other cases remain unsatisfactory.
Also, I've noticed that the _cpu_fallback_giou method seems to be detached from the computational graph, which is weird.

(Image 1 is the result from the original implementation, Image 2 is the worked around version)

Hi, I’ve been struggling with the same cal_my_giou problem. Would you mind sharing a bit more detail about how you modified the code (e.g., how you used the cal_giou method )?

Any pointers or snippets would be really helpful.

Hi, the function is under https://github.com/sunfanyunn/LayoutVLM/blob/main/third_party/Rotated_IoU/oriented_iou_loss.py#L75

Usage snippet:

    if ORIENTED_IOU_AVAILABLE:
        # Use the oriented IoU loss if available
        try:
            # Prefer cal_giou with (x, y, w, h, alpha) boxes
            giou_loss, iou = oriented_iou_loss.cal_giou(
                box1_tensor.to(device), box2_tensor.to(device).detach()
            )
        except Exception as e:
            warnings.warn(f"Error using oriented_iou_loss: {e}. Using simplified fallback.")
            giou_loss, iou = _cpu_fallback_giou(corners1, corners2, area1, area2, device)

Here's the docstring I wrote for DIoU. But it should be the same for GIoU.

    """calculate diou loss

    Args:
        box1: (B, N, 5),
        box2: (B, N, 5),
        last dimension: 
            - [x_center, y_center, width, height, angle (or alpha)]
    """