Closed Monoclinic closed 1 year ago
For lidar-only models, we found that rotation is useful in TTA. For MSMDFusion (lidar and image), we need to synchronize augmentation for both real 3D points and the virtual points from images, and the virtual points inevitably have some noise. We found that the rotation augmentation brings less improvement than the others (scaling/flip) in this case. So we simply combined scaling and flipping in our final TTA setting.
Got that. Thanks for your reply.
Hello, thanks for your great work. I found that your TTA (scaling/double flip) is different from other methods (rotation/double flip) on nuscenes leaderboard, could you please tell the reason?