search

edwardzhou130 / Panoptic-PolarNet

Implementation for Panoptic-PolarNet (CVPR 2021)
BSD 3-Clause "New" or "Revised" License
168 stars 31 forks source link

Do you know why the performance gap is so big between validation and test set? #11

Open xizaoqu opened 2 years ago

edwardzhou130 commented 2 years ago
  1. I used the same best pretrained model in the validation set to generate the test set result. I did not retrain the network with both train/val data or use TTA, both of which could improve the test set result by a lot (Used by many other methods in their test set submissions).
  2. The data distribution in semanticKITTI val and test sets are very big. You will find big result differences in most of the classes, especially those rare classes like motorcyclist.