Inquiry about Nuscenes Dataset Support and Code Updates in RegFormer

[Ruye-aa]

Thank you very much for your outstanding work!
I hope this message finds you well. I have been exploring the codebase of your RegFormer repository on GitHub, specifically looking into the handling of the Nuscenes dataset. I noticed that in HRegNet, there are distinct data preprocessing methods and parameters for both the Kitti and Nuscenes datasets.

My inquiry pertains to whether there is separate code in RegFormer specifically designed for preprocessing Nuscenes data. Could you please clarify if there are dedicated code for handling Nuscenes data within the RegFormer repository?

Furthermore, as I was examining the provided Google Drive link containing pre-trained models, I would like to know if there are models specifically trained on the Nuscenes dataset.

Lastly, I am interested in knowing if there are any plans for updating the codebase related to the Nuscenes dataset in the future. Understanding the timeline for potential updates would be greatly appreciated.

Thank you for your time and assistance. I am looking forward to your insights.

Best regards.

[liujiuming123]

Thanks for your acknowledgement of our work!
Basically, we follow the same data preparation method as HRegNet, using one frame and the 10-th frame after that to construct the input point cloud pairs. But we input all LiDAR points without voxelization process, which is different from HRegNet. In terms of data augmentation, we randomly rotate a certain angle to improve the generalization ability.
The models for KITTI and NuScenes are mostly the same one, apart from the projection parameters (as it depends on specific LiDAR sensor types).
We will release the model, parameters, and pretrained ckpt for NuScenes in the near future (maybe within the next two weeks).

Best regards.