IA-SSD
IA-SSD copied to clipboard
Effect of gt_sampling
Thanks for your sharing. When I train without gt sampling augmentation, I noticed the performance in moderate class drop from 83.4 to 77.11. As the gt sampling doesn't affect other methods like voxel r-cnn and pv_rcnn so obviously, can you explain this phenomenon?
Hi @Eaphan , very thanks for your following.
I think your finding is very illuminating. Although I have not tried this analysis experiment, I guess it is related to the representation ability of the model. After all, PV-RCNN and Voxel-RCNN are both two-stage frameworks, which will have more advantages in feature learning.
Hope this can help you.