Panoptic Segmentation

摘要: We propose and study a task we name panoptic segmentation (PS). Panoptic unifies the typically distinct tasks of semantic (assign class label to each pixel) instance (detect segment object instance). The proposed requires generating coherent scene that is rich complete, an important step toward real-world vision systems. While early work in computer addressed related image/scene parsing tasks, these are not currently popular, possibly due lack appropriate metrics or associated recognition challenges. To address this, novel quality (PQ) metric captures performance for all classes (stuff things) interpretable unified manner. Using metric, perform rigorous both human machine PS on three existing datasets, revealing interesting insights about task. aim our revive interest community more view image segmentation. For analysis up-to-date results, please check arXiv version paper: {\small\url{https://arxiv.org/abs/1801.00868}}.