search

fudan-zvg / DeepInteraction

[NeurIPS 2022] DeepInteraction: 3D Object Detection via Modality Interaction
MIT License
221 stars 16 forks source link

larger lidar backbone #7

Closed konyul closed 2 years ago

konyul commented 2 years ago

which large lidar backbone did you adopt compared to deepinteraction-base??

konyul commented 2 years ago

And do you need more epochs to converge deepinteraction-e compared to deepinteraction-base?? Thank you

Alexander0Yang commented 2 years ago

which large lidar backbone did you adopt compared to deepinteraction-base??

We still use VoxelNet in DeepInteraction-e, but enlarge some hyperparameters of the lidar backbone according to the voxel size.

And do you need more epochs to converge deepinteraction-e compared to deepinteraction-base??

No, we didn't tune the training schedule exhaustively for each model used in the ensemble version.

konyul commented 2 years ago

I found the hyperparameters of the larger lidar backbone in your paper, Thank you for the answering!!

Alexander0Yang commented 2 years ago

Nice.