search

TRAILab / PDV

Point Density-Aware Voxels for LiDAR 3D Object Detection (CVPR 2022)
Apache License 2.0
159 stars 30 forks source link

MeanOrientationVFE #10

Closed andy690 closed 2 years ago

andy690 commented 2 years ago

I found it in mean_vfe.py, I wonder what it does and how can I use him

jskhu commented 2 years ago

This is an artifact of from other experiments and PDV does not use this VFE method. However, to answer your question MeanOrientationVFE adds the ground truth bounding box orientation as an additional channel to the voxel feature input (xyz + intensity + elongation + orientation). One thing to keep in mind is that MeanOrientationVFE uses the ground truth labels during training: https://github.com/TRAILab/PDV/blob/f72f4f191f93d8a8ca301e46ebc69893ef6166e1/pcdet/models/backbones_3d/vfe/mean_vfe.py#L70-L72.

andy690 commented 2 years ago

Thank you for your reply. Perhaps assisting auxiliary network supervision to help establish the relationship between the orientation of voxels and ground truth can help us improve performance.

jskhu commented 2 years ago

Yup, that was one of the early ideas that we explored but didn't end up pursuing for this work.