Problem Reproducing the results on pedestrian and cyclist about openpcdet HOT 4 CLOSED

Update:

I also test the model PartA2_fc provided in the README, which gives a similar result.

from openpcdet.

This results is as expected for 3 classes of a the single model. Training pedestrian/cyclist with separate models and also modifying the config generally achieve better results. Also the performance of pedestrian/cyclist is not stable on KITTI dataset, and the released codes have been highly refactored and I mainly focus on car objects to tune the performance. You could also try to use a general classification loss instead of the IoU loss for tuning pedestrian, which generally achieve better results.

from openpcdet.

@sshaoshuai Thanks for your reply!

Could you please suggest how to modify the config file to reproduce the result in the paper except training in a separate model for ped/cyc and general classification loss?

from openpcdet.

Actually the cyclist is easier than pedestrian, I just try to train the cyclist with more times, and the performances between training cyclist with a single model or joint with car are similar. For the pedestrian, you could also try to upsample the BEV feature map to 4x for putting anchors as pedestrain has smaller size. Finally, I found the fact was that the anchor-free version (PartA2-free which generates proposals like PointRCNN) achieves better and more stable results on pedestrian/cyclist on KITTI dataset.

from openpcdet.