search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
915 stars 204 forks source link

Problem with some parameters in squeezeseg.yaml #73

Closed Zhen-ao closed 2 years ago

Zhen-ao commented 3 years ago

Thank you very much for your work But I’m not quite clear about some parameters in "squeezeseg.yaml"

  1. How is max_points=15000 obtained? How is it counted?
  2. Are img_means andimg_stds calculated by counting all points? If I understand correctly, does img_means_x calculate the sum of all points x and take the average?

best wishes

jbehley commented 3 years ago

Ooops, this is a rather old issue, therefore sorry that this slipped through.

  1. The maximum number of points considered is 150.000, which is well above the usual number of points in the KITTI dataset (110k -120k). I think the parameter is only relevant for architectures that need a downsampled point cloud to fit the data into the memory, such as PointNets, etc. Here, I think we used then random sampling.
  2. Here I have to refer to issue #43, where we already discussed that the real origin of the values is lost. But in principle, yes, one takes the range image and computes the mean over all valid entries in each channel and devices it by the number of valid entries.