Something wrong in computing 3D IoU

[yanjh97]

NOCS_CVPR2019/utils.py

Line 192 in 14dbce7

bbox_1_max = np.amax(bbox_3d_1, axis=0)

Hi, when I debugged method asymmetric_3d_iou(), I found that bbox_3d_1 shape is (3, 8).
Therefore, bbox_1_max shape is (8,), so are bbox_1_min, bbox_2_max, andbbox_2_min.
But I think their shapes should be (3,), which represented bounds of 3D boxes.

[hughw19]

There is nothing wrong. 8 is the eight vertices of a bounding box. Best, He

…

On Mar 21, 2021, at 8:51 PM, ColaerSugar ***@***.*** ***@***.***>> wrote: https://github.com/hughw19/NOCS_CVPR2019/blob/14dbce775c3c7c45bb7b19269bd53d68efb8f73f/utils.py#L192 <https://github.com/hughw19/NOCS_CVPR2019/blob/14dbce775c3c7c45bb7b19269bd53d68efb8f73f/utils.py#L192> Hi, when I debugged method asymmetric_3d_iou(), I found that bbox_3d_1 shape is (3, 8). Therefore, bbox_1_max shape is (8,), so are bbox_1_min, bbox_2_max, andbbox_2_min. But I think their shapes should be (3,), which represented bounds of 3D boxes. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#47>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEXSGH2RFWXZTQ5Z2CWYKDLTE25DRANCNFSM4ZSKARCQ>.

[yanjh97]

Hi, I did a test using method compute_3d_iou_new(). Results are as follows:

from utils import compute_3d_iou_new
import numpy as np
pose = np.eye(4, dtype=np.float32)
pose[:, 3] = np.array([1, 1, 1, 1])
print(compute_3d_iou_new(pose, pose, [1, 1, 1], [0.5, 0.5, 0.5], 1, 'laptop', 'laptop'))
#  def compute_3d_iou_new(RT_1, RT_2, scales_1, scales_2, handle_visibility, class_name_1, class_name_2): ...

• use bbox_1_max = np.amax(bbox_3d_1, axis=0) which output shape is (8,)

## output : nan

• use bbox_1_max = np.amax(bbox_3d_1, axis=1) which output shape is (3,)

## output : 0.125

It can be seen that when the output is (3,), the result is correct, but when the output is (8,), the result is strange.

[qq456cvb]

Hey guys, we've found that this piece of evaluation code is problematic not only because of this bug, but also inaccurate in computing IoU between arbitrary oriented boxes (this code is only accurate when the bounding box is axis aligned). To correct this, we borrow the code from Objectron, and you can find a bug-free version at https://github.com/qq456cvb/CPPF/blob/main/utils/util.py#L186.

[Van5SE]

Dear Author, does this problem means that all the following works calculated the wrong 3D IOU metrics and reported wrong results. Can we fix this problem by simply transpose the bbox_3d ?