search

dotchen / LAV

(CVPR 2022) A minimalist, mapless, end-to-end self-driving stack for joint perception, prediction, planning and control.
https://dotchen.github.io/LAV/
Apache License 2.0
397 stars 68 forks source link

Some question about the code #15

Open Watson52 opened 2 years ago

Watson52 commented 2 years ago

Hi Chen, thanks for the wonderful work. I would like to ask you some questions about the code in lav_agent.py.

  1. Does the npxs means the next target point about 50m away? Show as the green point in the video? image https://github.com/dotchen/LAV/blob/2d00c6276d1a72f45627b4b9b418cd2ccad0b994/team_code/lav_agent.py#L218
  2. In the inference phase i.e. using LAV agent, is it redundant to load the bev_model? Since I think the bev_planner (teacher) only used to generate soft labels for UniPlanner (student) to learn in the training time.
  3. One of the output of the lidar_model is named pred_heatmaps, but I think the corresponding output is the center location of other vehicles. Could you explain what a heatmapis? https://github.com/dotchen/LAV/blob/2d00c6276d1a72f45627b4b9b418cd2ccad0b994/team_code/lav_agent.py#L220-L224 https://github.com/dotchen/LAV/blob/2d00c6276d1a72f45627b4b9b418cd2ccad0b994/lav/models/lidar.py#L34-L44
dotchen commented 2 years ago
  1. Yes
  2. Absolutely correct. You can remove it from the code but you will need to add a strict=False when loading the weights.
  3. pred_heatmap is the predicted Gaussian "object centerness" map. It follows the CenterNet/CenterPoint [1,2] formulation

[1] Objects as Points, Zhou et al., 2019 [2] Center-based 3D Object Detection and Tracking, Yin et al., 2021