dvlab-research / VoxelNeXt

VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking (CVPR 2023)
https://arxiv.org/abs/2303.11301
Apache License 2.0
733 stars 64 forks source link

Large model for nuscenes #11

Closed c7huang closed 1 year ago

c7huang commented 1 year ago

Hi, thanks for publishing the code for your work!

I noticed the large model with 256 channels provides better results on Waymo, but the models for nuscenes and argoverse are not scaled up and use 128 channels. Is there a particular reason for this decision, or do you think the same scaling work on these datasets as well?

yukang2017 commented 1 year ago

Hi,

I only tried the large model on Waymo before, but not on nuScenes or Argoverse2 before. I think it should also be helpful.

Regards, Yukang Chen

c7huang commented 1 year ago

Thanks!