swc-17 / SparseDrive

SparseDrive: End-to-End Autonomous Driving via Sparse Scene Representation
MIT License
173 stars 9 forks source link

The augmentation is not applied for motion & planning #4

Open zhangyp15 opened 6 days ago

zhangyp15 commented 6 days ago

Hi, thanks for the great work and contributions.

When checking the data pipelines, I found that the motion & planning labels (though they are in LiDAR coordinates) are not transformed according to the augmentations. I wonder whether I missed some details or this is intended?

swc-17 commented 6 days ago

Data augmentations are only performed in image space rather than BEV space (rot3d_range is set to [0, 0]), so there is no need to transform motion & planning labels.

zhangyp15 commented 5 days ago

Thanks for your reply. I also found the baseline Sparse4D (for detection & tracking) performs the BEV augmentation with "rot3d_range": [-0.3925, 0.3925].

Why is it disabled for the end-to-end SparseDrive?

swc-17 commented 5 days ago

We found that data augmentation in BEV space seems to have negative effect for online mapping task.