Fix a small bug of rotate_iou_gpu: the iou between two same boxes can be 0.

[xuyongzhi]

A test demo is provided in test_nms_gpu.py

[traveller59]

This change can't get correct result when eval with same kitti gt annotations. you can try "rinter_cc" and "riou_cc" instead of "rotate_iou_gpu" to get correct result if this feature is really important. You can just set iou to 1.0 if their corners/boxes are close enough in devRotateIoU.
Use your change:

Car AP(Average Precision)@0.70, 0.70, 0.70:
bbox AP:100.00, 100.00, 100.00
bev  AP:6.19, 7.46, 9.82
3d   AP:6.19, 7.46, 9.82
aos  AP:100.00, 100.00, 100.00
Car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:100.00, 100.00, 100.00
bev  AP:6.19, 7.46, 9.82
3d   AP:6.19, 7.46, 9.82
aos  AP:100.00, 100.00, 100.00

Use c++ function:

Car AP(Average Precision)@0.70, 0.70, 0.70:
bbox AP:100.00, 100.00, 100.00
bev  AP:100.00, 100.00, 100.00
3d   AP:100.00, 100.00, 100.00
aos  AP:100.00, 100.00, 100.00
Car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:100.00, 100.00, 100.00
bev  AP:100.00, 100.00, 100.00
3d   AP:100.00, 100.00, 100.00
aos  AP:100.00, 100.00, 100.00

[xuyongzhi]

Thanks for you quick reply and detailed test! Just set as 1 may be a good choice, I will try it. Actually I also found some incorect cases with my solution today as well. In addition, may I ask which is the c++ function you mentioned?

[traveller59]

they are in second.core.box_np_ops, they call c++ functions from spconv/include/spconv/box_iou.h.