Skip to content

Conversation

xuyongzhi
Copy link

A test demo is provided in test_nms_gpu.py

0, due to point_in_quadrilateral. A test demo is provided in
test_nms_gpu.py.
@traveller59
Copy link
Owner

traveller59 commented Apr 16, 2019

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
Copy link
Author

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
Copy link
Owner

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

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants