MagicDrive About BEVFusion Data

Thank you for sharing the checkpoint for BevFusion. I want to reproduce CAM-ONLY results from Table 6. I generated data using the 272x736 checkpoint. I dropped 50% of the boxes both from the scene and the ground truth bboxes while generating.

I was wondering how did you fuse the new data to BevFusion? There is a data preparation step as the following: https://github.com/open-mmlab/mmdetection3d/blob/1.0/docs/en/datasets/nuscenes_det.md

How did you provide these for the generated data since we dropped some of the boxes? Can you give details about the dataloading process?

Thank you

Dec 17 '24 15:12 mrabiabrn

Hi, you can edit the metadata file and append the new infos with the generated images replacing the original ones.

Dec 18 '24 05:12 flymin

We have infos (which contain ground truth boxes, image paths, etc.) and database pickle files (which contain pointclouds). All I need to do is append my generated set's information to the infos files and not modify the database pickle files. Is this correct? But we are removing some of the objects, should we remove their point cloud data also? I am confused a bit. Thank you for your help.

Dec 19 '24 12:12 mrabiabrn

We did not modify the database pickle files. I think it should be safe at least for camera-only training. I am not sure about more details for this part.

Dec 20 '24 02:12 flymin

Thank you I figured out the data part by appending new information to train pickle file. I have several questions related to other parts:

I realized that for BevFusion data generation, you adjust the padding such that the generated images have black parts. Should we directly pass them to BevFusion?
Which CFG scale are you using?
Are you using the default seed for BevFusion?
In Table 6, when reporting 0.5x and 1x epochs, are you reporting the intermediate values from the full training (2x = 20 epochs) or are you training the model for 0.5x (5) epochs and 1x (10) epochs separately. The two differ due to learning rate scheduler.

Thank you in advance.

Dec 24 '24 07:12 mrabiabrn

This issue is stale because it has been open for 7 days with no activity. If you do not have any follow-ups, the issue will be closed soon.

Jan 01 '25 16:01 github-actions[bot]

May I ask how you managed to drop 50% of the bounding boxes during generation?

May 22 '25 13:05 mzsqw

This issue is stale because it has been open for 7 days with no activity. If you do not have any follow-ups, the issue will be closed soon.

May 30 '25 16:05 github-actions[bot]