Uncertainty-Aware Panoptic Segmentation

Reliable scene understanding is indispensable for modern autonomous systems. Current learning-based methods typically try to maximize their performance based on segmentation metrics that only consider the quality of segmentation. However, safe operation a system in real world it crucial uncertainty prediction as well. In this work, we introduce novel task uncertainty-aware panoptic segmentation, which aims predict per-pixel semantic and instance segmentations, together with estimates. We define two facilitate its quantitative analysis, Panoptic Quality (uPQ) Expected Calibration Error (pECE). further propose top-down Evidential Segmentation Network (EvPSNet) solve task. Our architecture employs simple yet effective fusion module leverages predicted uncertainties. Furthermore, provide several strong baselines combining state-of-the-art networks sampling-free estimation techniques. Extensive evaluations show our EvPSNet achieves new standard (PQ), well metrics.