How to implement point cloud supervision？

[CleanSeaSalt]

hello，Thanks much for such an excellent work!
I saw that the paper mentioned that the supervision results of adding point clouds were analyzed. I would like to ask for some ideas or methods to implement point cloud supervision. Thank you！

[yanglei18]

This codebase follows the same organizational structure as BEVDepth. Please refer to the specific implementation of BEVDepth for details.

[CleanSeaSalt]

OK
I used “/scripts/data_converter/visual_tools.py ”to generate some ground truth, but there are some obvious errors, such as ：


What step went wrong?Thanks!

[CleanSeaSalt]

Hello, I would like to ask what the histogram drawn in section 3.1 of the paper, the horizontal axis range [-1,2], means? Does it mean the height of the horizontal plane where the radar is located as the reference? That is to say converted to [5,7]?

[CleanSeaSalt]

Hello, I have done a lot of experiments with reference to the supervision method of BEVDepth (point clouds are projected onto pictures, through minimum pooling and one hot encoding), but they cannot exceed the unsupervised results. I would like to ask about the one hot encoding part. Isn't it wrong that a discrete distribution should be set?