Usage on KITTI about rotated_iou HOT 7 OPEN

Hi, thanks for the great work for IoU Loss for 2D/3D Object Detection.

i was wondering if there is the implementation for 3d object detection. i can see your work focus on the 2d calculation, right?

from rotated_iou.

Hi, thanks for your great work. I want to use these IoU losses on KITTI and evaluate the model performance. How should I do, and how to write the transform scrip? Thanks in advance!

I'm not sure which coordinate system you are using. But what you need is basically a coordinate transformation. My code uses a right-hand system with x pointing to right, y to front and z to up. The rotation around z-axis starts from x-axis. In most cases, you only need to swap the axis and multiply -1 to some coordinates.

from rotated_iou.

Hi, thanks for the great work for IoU Loss for 2D/3D Object Detection.

i was wondering if there is the implementation for 3d object detection. i can see your work focus on the 2d calculation, right?

There are functions for 3D cases. Please check cal_iou_3d, cal_giou_3d and cal_diou_3d in oriented_iou_loss.py. Since most 3D detection task only considers the rotation around vertical axis, these 3D functions are actually extended from the 2D counterpart.

from rotated_iou.

Hi, thanks for your great work. I want to use these IoU losses on KITTI and evaluate the model performance. How should I do, and how to write the transform scrip? Thanks in advance!

I'm not sure which coordinate system you are using. But what you need is basically a coordinate transformation. My code uses a right-hand system with x pointing to right, y to front and z to up. The rotation around z-axis starts from x-axis. In most cases, you only need to swap the axis and multiply -1 to some coordinates.

Thanks for reply. I am also confused about the coordinate that the KITTI dataset uses, and directly use the official label files.
In this case, what should I do to calculate the 3D IoU/GIoU loss as you provided?

from rotated_iou.

@ChunmianLin You can transform either the label or the prediction. The transformation depends on the coordinate system of KITTI. I'm not familiar with KITTI so cannot provide precise suggestions.

from rotated_iou.

Hi, thanks for the excellent work for IoU Loss for 2D/3D Object Detection.
I was wondering if there is an implementation for 3d object detection. I can see your work focus on the 2d calculation, right?

There are functions for 3D cases. Please check cal_iou_3d, cal_giou_3d and cal_diou_3d in oriented_iou_loss.py. Since most 3D detection task only considers the rotation around vertical axis, these 3D functions are actually extended from the 2D counterpart.

Thank you for your quick reply.

from rotated_iou.

@ChunmianLin You can transform either the label or the prediction. The transformation depends on the coordinate system of KITTI. I'm not familiar with KITTI so cannot provide precise suggestions.

Thanks. In your work, which dataset does you pick? Maybe I could find some idea from your advice.

from rotated_iou.