About the VoVnet-99 checkpoint and config

[zw615]

Hi, thanks for the great work!

I wonder if you could release the VoVnet-99 version checkpoint and config, so that the test set results can be reproduced?

Also, I wonder what pre-trained checkpoint for VoVnet-99 did you use? Was it first pre-trained on the COCO dataset for 2D Detection, and then pre-trained on the DDAM-15M dataset for depth prediction? Did you freeze any layer/stage in VoVnet-99 when training for 3D Detection?

Thanks a lot!

[swtju14]

Hi, @zeyuwang615 . We adopt DD3D pretrained VoVNet (V2-99) as backbone network for images. No freeze when training.

[zw615]

Does 'DD3D pretrained VoVNet (V2-99)' means the backbone is pre-trained sequentially on COCO and DDAM-15M as described in the paper?

More importantly, is there any checkpoint for such DD3D pretrained VoVNet (V2-99) available? I have managed to find two, but not sure which is the right one (or neither is): here in dd3d codebase

or dded_det_final.pth here in detr3d codebase.

Thanks a lot!

[swtju14]

dded_det_final.pth here in detr3d codebase.

[thom966]

Could you please provide the vovnet.py file that you used when loading pre-trained weights?

[swtju14]

We use vovnet.py file like this