search

open-mmlab / mmdetection3d

OpenMMLab's next-generation platform for general 3D object detection.
https://mmdetection3d.readthedocs.io/en/latest/
Apache License 2.0
5.3k stars 1.54k forks source link

Point cloud range info #644

Closed selmadeac closed 3 years ago

selmadeac commented 3 years ago

Hi! I noticed the point cloud range variable is different for SECOND and for POINTPILLARS when looking on kitti config files. For example, for SECOND it is: [0, -40, -3, 70.4, 40, 1] and for pointpillars: [0, -39.68, -3, 69.12, 39.68, 1]. If I try to change it and put a random number, the network does not train well. I get 0% mAP on all classes. Is it because the values are computed based on the given voxel_size? And why is the max_voxels=(16000, 40000), and not just an integer? What does 40000 stand for?

Wuziyi616 commented 3 years ago

Of course the random number won't work, because this limits the point cloud you will see. Those numbers are taken from their papers. They are basically some empirical number and not related to voxel_size.

As for the tuple input of max_voxels, it seems from here that, the first value is used in training mode and the second value is used in testing mode. Please note that we will drop the remaining points if voxels number goes beyond max_voxels by voxelizing point clouds by voxel_size. So using a larger max_voxels during testing can keep more points and thus get better performance. You can refer to their paper for more details.