Question with configs_pgd backbone_img and Deformable Convolution

[MinHsuan123]

Hello, I saw a section of dcn=dict(type='DCNv2', deform_groups=1, fallback_on_stride=False) in the backbone_img in configs_rqadiant_pgd. Does this mean that the use of Deformable Convolution is prohibited? If I want to try adding Deformable Convolution to your code, is it possible? Thank you

[longyunf]

To use pretrained weights for camera backbone, I follow the standard model configuration in PGD, i.e. https://github.com/open-mmlab/mmdetection3d/blob/5c0613be29bd2e51771ec5e046d89ba3089887c7/configs/pgd/pgd_r101-caffe_fpn_head-gn_16xb2-1x_nus-mono3d.py#L14C1-L14C75.

[MinHsuan123]

To use pretrained weights for camera backbone, I follow the standard model configuration in PGD, i.e. https://github.com/open-mmlab/mmdetection3d/blob/5c0613be29bd2e51771ec5e046d89ba3089887c7/configs/pgd/pgd_r101-caffe_fpn_head-gn_16xb2-1x_nus-mono3d.py#L14C1-L14C75.

Hello~ So there is no way to add Deformable Convolution to your camera branch or radar bracng? Thanks

[longyunf]

I believe it is using deformable convolution.

[MinHsuan123]

I believe it is using deformable convolution.

Are you referring to camera branch or radar branch? Or both? Can you explain where the settings are in the code? Because I found that the PGD paper didn’t mention that they had used deformable convolution. Thanks

[longyunf]

PGD is using deformable convolution (https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/configs_radiant_pgd.py#L27-L28). Radar backbone (https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/configs_radiant_pgd.py#L29) is not. You can modify arguments dcn and stage_with_dcn in resnet for radar backbone to use it. (https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/my_model/resnet.py#L373-L374)

[MinHsuan123]

PGD is using deformable convolution (

https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/configs_radiant_pgd.py#L27-L28

). Radar backbone ( https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/configs_radiant_pgd.py#L29

) is not. You can modify arguments dcn and stage_with_dcn in resnet for radar backbone to use it. ( https://github.com/longyunf/radiant/blob/cf5355396d42ef17940e29ef8f9e3cabfd8035c3/lib/my_model/resnet.py#L373-L374

)

Thank you for your help. I also want to ask, if I want to change the model of the camera branch, can I just change the"--path_checkpoint_pgd" parameter directly? Or do config and network also need to be changed accordingly? Thank you