the world2lidar transformation matrix compute

[Y-vic]

Hi, guys! I have found that there are two ways in the current repository (like file bench2drive_drivetransformer_dataset.py) to compute the world2lidar transformation matrix, but they are not equal.
1st way:

lidar2ego = cur_frame['sensors']['LIDAR_TOP']['lidar2ego']
ego2world = np.eye(4)
ego2world[0:2,3] = cur_frame['ego_translation'][0:2]
ego2world[0:3,0:3] = Quaternion(axis=[0, 0, 1], radians=cur_frame['ego_yaw']).rotation_matrix
lidar2world = ego2world @ lidar2ego
world2lidar_lidar_cur = self.invert_pose(lidar2world)

2nd way:

world2lidar_lidar_cur=cur_frame['sensors']['LIDAR_TOP']['world2lidar']

However, there is some error between both, which I think is a little large(like 0.6m). Below is an example of the minus result I computed for one frame:

Can you explain more about this?

[zhiyuanzzz]

Hi @Y-vic . Actually, the meaning of world2lidar transformation caltulated by two methods are different.

• The 1st way gets the estimated pose from GPS sensor of ego with some noise simulated in CARLA, and 0.6m is in normal range. This estimated pose is than used to calculate past trajectory of ego car as input of the model. This data form is available in closed-loop evaluation. We use it in training to align with the evaluation.

• The 2nd way gets the actual pose from CARLA APIs. This pose is than used to calculate future trajectories as GT. It is unavailable in closed-loop evaluation.