Use the point cloud frame generated by Carla for prediction and training

[S-kewen]

Hi Shaoshuai:

First of all thank you very much for your contribution, this code is very powerful.

I am recently trying to use PointRCNN to predict the lidar point cloud frame generated by Carla, but the results are not good. I guess the frame generated by Carla is different from KITTI (I used KITTI to reproduce the results successfully)

So I tried to collect the dataset (KITTI-format) in Carla for PointRCNN training (rpn+rcnn), but I trained according to the default parameters and found that the result is still not ideal, I would like to ask the point cloud frame Does intensity matter? Because the intensity in Carla only refers to the distance, it is not real.

Here is the result of the Carla dataset:
Car AP@0.70, 0.70, 0.70:
bbox AP:34.6747, 31.2628, 31.0747
bev AP:16.8469, 15.9205, 17.7584
3d AP:13.8553, 12.0301, 14.5156
aos AP:25.04, 23.67, 23.37
Car AP@0.70, 0.50, 0.50:
bbox AP:34.6747, 31.2628, 31.0747
bev AP:51.1212, 42.6968, 43.5934
3d AP:48.2478, 35.7272, 36.8692
aos AP:25.04, 23.67, 23.37

Looking forward to your reply, thanks!