search

edwardzhou130 / PolarSeg

Implementation for PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation (CVPR 2020)
BSD 3-Clause "New" or "Revised" License
380 stars 80 forks source link

Train on a dataset and test on another #40

Closed SofianeB-03 closed 3 years ago

SofianeB-03 commented 3 years ago

Hello, thank you very much for your great work ! I have a question: when we test the pre-trained model on a new database, should we modify the axes of the point cloud such that they are the same as the training database ? For example if I test on Nuscenes the pre-trained PolarSeg on Semantickitti, should I modify the axes of Nuscenes according those of SemanticKitti ?

edwardzhou130 commented 3 years ago

It's not necessary to modify the axis of the point cloud if the only difference between your dataset and KITTI is the rotation of the axis. The bigger problem is the number of class and class definitions. You need to carefully match the class in your dataset to the KITTI dataset to get a meaningful prediction. And some other differences like the number of points/ range also could have a big influence on the result.

SofianeB-03 commented 3 years ago

Thanks for your reply ! A last question: is it because the data has been augmented (translation + rotation) during training that we do not need to modify the axes of the point cloud?

edwardzhou130 commented 3 years ago

Yes, we did the rotation and flip augmentation in training.

SofianeB-03 commented 3 years ago

Thank you !