search

V2AI / Det3D

World's first general purpose 3D object detection codebse.
https://arxiv.org/abs/1908.09492
Apache License 2.0
1.48k stars 299 forks source link

Shuffling of KITTI validation examples #115

Closed afroowoo closed 4 years ago

afroowoo commented 4 years ago

Hi, in the pointpillars kitti config file it seems that the points are not shuffled during validation:

https://github.com/poodarchu/Det3D/blob/255c5937014cf65d5eac3c1066a84d7c719b1ef3/examples/point_pillars/configs/kitti_point_pillars_mghead_syncbn.py#L151-L153

I believe the KITTI pointclouds are organized from far to near, so during voxelization the first 12000 voxels might be filled without exploring the nearby points.

So the question is, is shuffling of KITTI val necessary?

tianweiy commented 4 years ago

i think the idea is that you want to get some consistent results for validation. Shuffling will create randomness. Also for kitti 12000 pillars should be fine to cover most of the points.

poodarchu commented 4 years ago

Hi, in the pointpillars kitti config file it seems that the points are not shuffled during validation:

https://github.com/poodarchu/Det3D/blob/255c5937014cf65d5eac3c1066a84d7c719b1ef3/examples/point_pillars/configs/kitti_point_pillars_mghead_syncbn.py#L151-L153

I believe the KITTI pointclouds are organized from far to near, so during voxelization the first 12000 voxels might be filled without exploring the nearby points.

So the question is, is shuffling of KITTI val necessary?

It should be off, and this doesn’t affect the results