Semantic Segmentation weight initialization


I’m working on a semantic segmentation task in point clouds, I have 9 classes, labeled from 0 to 8. when 0 is unlabeled points.

The classes are widely varied with the number of points, vegetation and ground are around 70% of the point cloud, and fence and light poles account for very few points. I understand that this affects the training process, so we should initialize the weights with respect to the proportion of the classes.
I’d like to have more explanation regarding this topic.
Thanks very much.

Still have gotten no answer, I suppose it causes the IoU metric to go down while training.

I don’t think so this a good idea, a better idea would be ground plane augmentation.

What do you exactly mean by ground plane augmentation? never heard of it. Thanks!