I found that the CenterFormer is categorized as the dense detector in Table 1. But as far as I know, the CenterFormer uses 500/1000 queries produced by the topK method to detect objects on BEV features, so I think it should be a sparse method. I don't know if there is a mistake in my understanding, how do you define the sparse detectors in your paper?

The command is "efg_run task", but how could I run the ConQueR code?
Thank you.

Hi, I found that there is no config of large backbone and 4x resolution about ConQueR in the repo.
Can you share the config? Thanks.

Hi, thanks for your great work on ConQueR.
Where can I find the loss calculation section of Query Contrast? I think I may have missed something, please guide me.

cd playground/path/to/experiment/directory This path does not exist. May I ask how to start training？

I am debugging at '/EFG/cli/main. py' and I am missing the 'EFG/cli/config. yaml' file. May I know how to obtain this file？
In addition, can I inquire about how many GPUs this method uses.

Hello! What's ConQueR's number of parameters?

This is a great work, thanks for sharing the code. The paper has always been experimentally verified on the nuScenes data set. I wonder if you can provide the relevant code of this method in the nuScenes data set, thank you.

Congratulation on the acceptance of ConQueR. I'm very interested in your work. When will you release your code?

Thanks for sharing your concise and elegant framework.

When I try to understand the code of VoxelDETR, I found some code may be duplicate. It's unnecessary to execute the forward operation of decoder.detection_head twice.

https://github.com/poodarchu/EFG/blob/dedd7375eb02ef163bc04343973f8c35fc0e66f8/projects/detection.3d/waymo/conquer/VoxelDETR.waymo.res18.p3.box_only_with_3cat.bs6.epoch6/voxel_detr.py#L135