Hello, I made a customized dataset with 50+ point cloud files and converted it to kitti format. However, the training results of pointpillars on it are weird.

As is shown above, both bbox and aos are 0 accuracy, which is pretty strange. I visualized the results of several samples and found out that the predicting results are roughly acceptable so that neither 2D nor direction evaluation result should be 0 accuracy.

Does any one know how to solve this problem? Thanks!