search

open-mmlab / mmdetection3d

OpenMMLab's next-generation platform for general 3D object detection.
https://mmdetection3d.readthedocs.io/en/latest/
Apache License 2.0
5.32k stars 1.54k forks source link

FCOS3D vs PGD #2482

Open dadaligoudan opened 1 year ago

dadaligoudan commented 1 year ago

Hi， I trained mono FCOS3D and PGD model on my own dataset, but PGD performs more worse than FCOS3D both on bbox AP11 and the final detections. Could someone help me figure it out. The config models of FCOS3D and PGD are barely modified. FCOS3D: Overall AP11@easy, moderate, hard: bbox AP11:56.9753, 54.4123, 54.4123 bev AP11:21.2570, 18.2347, 18.2347 3d AP11:18.2296, 15.2323, 15.2323 aos AP11:32.04, 30.28, 30.28 PGD: Overall AP11@easy, moderate, hard: bbox AP11:41.9162, 36.5487, 36.5487 bev AP11:22.0470, 18.5532, 18.5532 3d AP11:19.5883, 17.4104, 17.4104 aos AP11:24.27, 21.20, 21.20 I want to know what reason that caused this problem. Thanks a lot.

JingweiZhang12 commented 1 year ago

Maybe some hyper-params about PGD are not proper on your own dataset? @Tai-Wang

dadaligoudan commented 1 year ago

I set almost all the hyper-params the same when training FCOS3D and PGD, may be that is not the reason.

dadaligoudan commented 1 year ago

Hi，when i trained FCOS3D model on my dataset, I changed GN to BN in FCOSMono3DHead for model deployment, but the training result was so bad. The bbox AP11 degress from 60 to 30, and the 3d AP 40 degress from 20 to 7. Therefore, the FCOS3D model trained with BN can't be used. This phenomenon cannot be explained theoretically。What should i do to use BN in FCOSMono3DHead to get a good training result? Could you please help me with it? Thanks. @Tai-Wang