About coordinate system transformation, ego system and Lidar system

[lilkeker]

Hello Dr. Huang, Bevdet is an excellent work! Now I want to implement ego's bevfusion work on the basis of bevdet, but I have encountered some problems:

1. About camego and lidarego: I found camego and lidarego here, and in the experiment lidarego The conversion process of ar2cam is lidar-->lidarego-->global-->camego-->cam, I want to know if lidarego and camego are different "ego" systems, if I want to realize lidar system-->ego system transformation of point cloud data, which one should I use?

2. About LiDARInstance3DBoxes, origin=(0.5, 0.5, 0.5)in bevdet & origin=(0.5, 0.5, 0) in bevfusion: I found a difference from bevfusion in here, the same location in bevfusion, ,I understand that this place is to change the center of the box, but I don’t quite understand why bevdet and bevfusion are different in this place, Is it because Bevdet is in the ego system , and bevfusion is in the lidar system?

3. About implementing the Point cloud detection model in the ego system of bevdet: I use the lidar2lidarego conversion process in PointToMultiViewDepth, transfer the points to the ego system for training, and use The Gt of the ego system of ann_infos in bevdetv2-nuscenes_infos_train.pkl, but the model I trained has not been able to reach the mAP and NDS of the lidar system training of the model（in ego system:mAP=0.37,in lidar system:mAP=0.65）, Is there something wrong with the conversion of my lidar2ego or something else? BTW, I have done related visualization, point and gt_boxs are aligned, the following is mine Visualization results and associated code: points_lidar = results['points'] points_inlidargo = points_lidar.tensor[:, :3].matmul( lidar2lidarego[:3, :3].T) + lidar2lidarego[:3, 3].unsqueeze(0)

Looking forward to your reply and answers, it will be very helpful to me!