Question about per-sample ECE #3
Description
Hi, thanks for the awesome work.
I noticed that you adapted the ECE metric for semantic segmentation. I'd like to know the specific reason for this and whether you took into account the different class distributions, as NuScenes and SemanticKitti are unbalanced datasets. Since this metric is calculated per point cloud, it does have a little bit of "regularization" effect to prevent overreliance on the majority class. However, this also strongly depends on the exact distribution of each point cloud.
Have you considered variants such as class-wise ECE? or maybe even some variant of Adaptive Calibration Error(ACE)