Bug in nuscenes_dataset.py

[cvpr2023-3d-occupancy-prediction]

Bug in nuscenes_dataset.py

[tongwwt]

Please check again the lin208-212 in https://github.com/CVPR2023-3D-Occupancy-Prediction/CVPR2023-3D-Occupancy-Prediction/blob/HEAD/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py#L208-L212
The original translation and rotation in can bus has been replaced with ego2global transformation.

• rotation = Quaternion(input_dict['ego2global_rotation'])
• translation = input_dict['ego2global_translation']
• can_bus = input_dict['can_bus']
• can_bus[:3] = translation
• can_bus[3:7] = rotation

[tongwwt]

Hi, can you describe your question in detail? The rotation matrix is equivalent to the quaternion when defining the transformation matrix.

[Isaac-xie]

in nuscene_dataset.py, line 212
can_bus[3:7] = rotation
I guess the rotation (0.394-0.010i-0.003j+0.919k) is to be assigned to can_bus[3:7], but after the assignment, can_bus[3:7]=[0.394, 0.394, 0.394, 0.394]. why can_bus[3:7] != [0.394, - 0.010, -0.003, 0.919]?

[Isaac-xie]

@tongwwt
hi，I have not modified any code in nuscenes_dataset.py. If you run it without the above problem, is there a problem with the version of Quaternion？

Here is my debug screen, please note can_bus[3:7]

I think line 212 should be changed to can_bus[3:7] = input_dict['ego2global_rotation']

[tongwwt]

Hi, I check the code. The can_bus[3:7] only keeps the information of roatation angle which can represent the quaternion, causing the problem you provides. This type of definition does not affect the model performance.