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

GT-AUG func problems #146

Closed majaya closed 3 years ago

majaya commented 3 years ago

I'd like to ask some questions,when I train model using Det3D project, I find that GT-AUG strategy is not exactly precise sometimes. the GT-AUG func in Det3D project only take account collision with the origin gt boxes, but the boxes that we get from gt_database sometimes will embedded in the wall, bush or other entity in current point cloud frame, also, some boxes will get off the ground and hanging in the air. Is it necessary to consider avoiding this situation. Does this affect training?

poodarchu commented 3 years ago

According to my experiments, this problem does indeed exist, but it has minor effect on the final results. I think it might due to the gain of GT-AUG w.r.t. number of instances overweight the side-effect caused by the collision problem. To avoid this collision, the segmentation map is required, or rough approximation of the wall are required.