Open kyrie-louy opened 2 years ago
I am sorry about this situation. I have tried several different 2D detection methods through mmdet, while they works well. I've never encountered a situation like this before. You may verify that your 3D detection results are produced by OpenPCDet and that your 2D detection results do not go through NMS. I'm afraid I can't make any additional suggestions. I hope you can resolve this issue.
Hello, i tried this technique to compute clocs and the results are really bad, Can you help me with this?
2023-02-06 14:49:50,138 INFO Car AP@0.70, 0.70, 0.70: bbox AP:4.2648, 4.2198, 3.2463 bev AP:4.1331, 4.0580, 3.1242 3d AP:4.0527, 3.8060, 2.9328 aos AP:4.26, 4.18, 3.21 Car AP_R40@0.70, 0.70, 0.70: bbox AP:4.4958, 4.3517, 3.2733 bev AP:4.1676, 4.1848, 3.1503 3d AP:4.0865, 3.6632, 2.6884 aos AP:4.49, 4.31, 3.24 Car AP@0.70, 0.50, 0.50: bbox AP:4.2648, 4.2198, 3.2463 bev AP:4.2682, 4.2643, 3.2814 3d AP:4.2680, 4.2537, 3.2734 aos AP:4.26, 4.18, 3.21 Car AP_R40@0.70, 0.50, 0.50: bbox AP:4.4958, 4.3517, 3.2733 bev AP:4.4994, 4.3975, 3.3087 3d AP:4.4992, 4.3866, 3.3007 aos AP:4.49, 4.31, 3.24
Hi @Laiqingsi, thanks for your great work. It helps me a lot!
However, I found that the AP improvement by fusion is very low (even degradation in some cases) compared to original CLOCs implementation. For example, when I train the fusion model with my own 2d detection results from Retinanet in Detectron2, it doesn't gain anything from fusion. Oppositely, I can get a 3%+ improvement in the original implementation.
It's quite strange as I can't find an obvious difference between your code and the original implementation. I wonder that have you encountered such an issue when evaluating different 2d predictions? What may be the possible reason?
Notes: I've also tried to increase the 3D prediction number to 70400 for the second model, but got a similar observation.