dennis-park-TRI
commented
2 years ago Hello @JohanCao , thanks for checkout out our work!
Here's a DD3D model with V2-99 backbone that is pre-trained only on COCO: https://tri-ml-public.s3.amazonaws.com/github/dd3d/pretrained/coco_pretrained_v99_1cinuqqf-20210513_071634_model_final.pth But please use it at your risk, because this model might be slightly outdated, and the keys of the state_dict might not perfectly match the one in the release model. If it works, it should give 18.45% BEV AP on Car Moderate on KITTI val. If it does not, you might have to remap the keys in the checkpoint to match the DD3D state_dict in the released code. In the early part of training log, it prints what weights are used and unused, so you will know the weights are correctly loaded. If there are more demands for this checkpoint, I will work on the remapping and officially release it. Another option is to actually train the model yourself, starting from the COCO pre-trained V2-99 backbone, which is available from here: https://dl.dropbox.com/s/v64mknwzfpmfcdh/faster_V_99_eSE_ms_3x.pth. (officially released here: https://github.com/youngwanLEE/vovnet-detectron2). You will have to do a bit of key remapping too.
And to answer your question, yes, the released models are pre-trained on both COCO and DDAD15M:
JohanCao
Hi TRI team,
Thanks for sharing this exellent work! I have lately been running this code and want to run some experiments from the scratch. Could you please share your DD3D model pretrained ONLY on the COCO dataset (without depth pretraining)?
And also, just want to be clear, the model you have provided are pretrained on COCO AND DDAD15M, right?
Thank you and I am looking forward to your reply!
Regards, Johan