About the gt_bboxes_3d coordinates? Ego or Lidar?

[YTEP-ZHI]

BEVerse/projects/mmdet3d_plugin/datasets/mtl_nuscenes_dataset_ego.py

Lines 462 to 469 in 5c6f4f7

	# convert 3DBoxes from LiDAR frame to Ego frame
	lidar2ego_translation = info['lidar2ego_translation']
	lidar2ego_rotation = info['lidar2ego_rotation']
	lidar2ego_rotation = pyquaternion.Quaternion(
	lidar2ego_rotation).rotation_matrix
	gt_bboxes_3d.rotate(lidar2ego_rotation.T)
	gt_bboxes_3d.translate(lidar2ego_translation)

I was wondering why you translate the coordinates systems of gt_bboxes from Lidar to Ego? Why not directly use Lidar coordinates?

[YTEP-ZHI]

@zhangyp15 It seems that this transformation is not needed in other 3d detection projects like BEVFormer.

[zyrant]

I think it is perhaps a need for multi frames.

[alchemz]

@zyrant I think you are right, in the paper, the author mentioned ego-motion alignment performed after features extraction of multi frames.

@YTEP-ZHI Is lidar coordinate equivalent to world coordinate?