Maybe bug in evaluation about omni3d

[garrickbrazil]

Hello! Thank you for submitting this issue. I'll look into it over the next few days. I wonder if it's related to a caching issue?

Can you elaborate on what you mean by "delete all 3D detection results"? My understanding is that the code shouldn't evaluate at all if the 3D fields have been deleted. (also, if you have any cached json files for the kitti_val detections to help my inspection, feel free to share, otherwise I'll reproduce.)

[Lizhuoling]

I simply delete the values corresponding to keys ['pred_bbox3D', 'pred_center_cam', 'pred_center_2D', 'pred_dimensions', 'pred_pose'] in the results. It is realized by the following code line109~line111 in rcnn3d.py:

The evaluation can still run normally. It seems because the 2D object detection results are provided, and all missed 3D object detection results (the values of deleted keys) will be padded as 1 automatically.

[garrickbrazil]

Thanks for the details! This is unexpected behavior in the evaluation which we've fixed by adding coplanar and nonzero checks for the cuboids, which may otherwise cause invalid (infinite) 3D IoUs. The issue should be fixed in the latest commit. Feel free to close if you find the issue is resolved as well.

[Lizhuoling]

Thanks for your effort! My problem has been addressed.

[star9988rr]

Hi, @Lizhuoling, can you reproduce the results of KITTI? What hyper-parameters do you use?