Estimating label quality and errors in semantic segmentation data via any model
The soft-minimum of the model-estimated likelihoods of each pixel’s annotated class – that is particularly effective to identify images that are mislabeled, across multiple types of annotation error