Question about the configuration files and model code for outdoor object detection on the Waymo dataset

[CheeryLeeyy]

First of all, I greatly admire your series of excellent works, especially the latest PointTransformerV3. I am particularly interested in the "Outdoor object detection on Waymo" section and would like to experiment with it to further understand the ingenious point cloud serialization process. However, during the debugging process, I encountered a few issues:

1. I couldn't find the configuration file for "Outdoor object detection on Waymo." Could you please tell me its path?
2. This method is based on CenterPoint for object detection, but I am unclear about the model's output (PointTransformerV3 in pointcept/models/point_transformer_v3/point_transformer_v3m1_base.py). How are the parameters needed for object detection, such as cls_preds and box_preds, generated and structured?
3. I couldn't find the loss function used for training CenterPoint.

Additionally, I am curious whether the serialized point cloud is stored as a token sequence, such as an array of shape (b, l, d). In scenes with different numbers of vehicles, how is the sequence length controlled, and how are varying numbers of bounding boxes (bbox) generated?

Thank you very much for your patience and guidance in answering these questions.

[Gofinge]

Hey, according to the code for object detection. It was based on openpcdet, and I didn't have time to clean the code. Drop me an email for the raw code and I might upload it to another branch of the PTv3 repo later.

According to how we handle the scenes with different numbers of vehicles (do you mean shape?), check here: https://github.com/Pointcept/Pointcept?tab=readme-ov-file#offset