Initialization weight of PolarFormer-Vov.

[Daniel-xsy]

Sorry to bother you again. Can you please provide the initialization weight of the PolarFormer-Vov since it's initialized from DD3D as mentioned in the paper? I did not find it in this codebase. Thanks so much!

[yanqinJiang]

Hi, the vovnet-pretrained model is downloaded from DETR3D repo, named dd3d_det_final.pth there.

[Daniel-xsy]

@yanqinJiang Thanks for your information.

[Daniel-xsy]

Hi @yanqinJiang

I noticed in https://github.com/megvii-research/PETR/issues/75#issuecomment-1323666167, it mentioned that the dd3d_det_final.pth is first pretrained on DDAD15M dataset and further trained on nuScenes trainval set with DD3D. So I'm wondering if is it reasonable to evaluate this version on nuScenes validation set due to information leakage (nuScenes validation set is used during DD3D pre-train).

We are working on benchmarking ood robustness of camera-only detectors, and PolarFormer-Vov shows superior performance. I'm wondering if it is caused by the reason mentioned above. If you are interested in the results, please kindly refer to RoboDet.

[Daniel-xsy]

The results shown in this codebase of PolarFormer-Vov on nuScenes validation set also largely surpass the ResNet version.

[yanqinJiang]

I didn't notice that dd3d_det_final.pth was trained on nuScenes trainval set. Thanks for pointing out my mistake. Of course, it is unfair to evaluate PolarFormer-Vov initialized from dd3d_det_final.pth on validation set due to information leak. I will delete this result. Thanks again. For fair comparison, you can train a PolarFormer-Vov initialized from fcos3d_vovnet_imgbackbone-remapped.pth as mentioned in PETR repo.

[Daniel-xsy]

Thanks for the update.