The original preprocessing pipelines (for the ADE20k dataset for instance) can be found here.