Uncertainty-Aware Panoptic Segmentation

The ternary difficulty levels assigned to each pixel during the annotation of MUSES enable a novel task: uncertainty-aware panoptic segmen- tation. In this task, a panoptic segmentation model is compensated for errors in difficult image regions if it predicts the difficulty level correctly. To incorporate this idea into evaluation, we introduce the uncertainty-aware panoptic quality (UPQ) metric, an extension of panoptic quality.