CenterPoint+PointPainting on nuScenes Val dataset

[buaacyw]

Hi! Thanks for your paper MVP. But I have a question about the result of CenterPoint+PointPainting on nuScenes Val dataset. As you mentioned in the paper, it reaches 66.5 overall mAP. But my result is only 65.4. I train the model by just throwing away those virtual points since you have already painted those real points on the masks. All other configurations are the same with https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale_virtual.py

Could you please tell me the reason? Thanks!

[buaacyw]

[tianweiy]

Hi, in the table, we also do the same second stage for point painting. For single stage, your number seems right

[buaacyw]

Thanks for your quick reply! As for the second stage, I find in https://github.com/tianweiy/CenterPoint/blob/d1e1bd5b08b2b9816ee38bbe2d036db1e5df340a/configs/mvp/nusc_two_stage_base_with_virtual.py#L278 that you have only trained 6 epochs. Does that mean you first only train one stage for 20 epochs and then resume from those parameters and train the second stage for 6 epochs?

[tianweiy]

Yes, set the weight trained using first stage at https://github.com/tianweiy/CenterPoint/blob/d1e1bd5b08b2b9816ee38bbe2d036db1e5df340a/configs/mvp/nusc_two_stage_base_with_virtual.py#L27

[buaacyw]

Thanks! And I found that you haven't removed those raw points which are within those masks and are painted. For MVP, I think you keep those raw points for split voxelization. But just for CenterPoint+PointPainting, why not remove those raw points away?

[tianweiy]

This also works. I just keep the comparison the same as our approach in table 4.

[buaacyw]

Thanks!