Panoptic Segmentation
Overview
Panoptic segmentation combines semantic segmentation of background-like regions (“stuff”) with instance segmentation of countable objects (“things”), giving a unified account of a scene. Common approaches use query-based transformer decoders, unified-kernel models, and temporal or geometric association, with extensions to open-vocabulary recognition, 3D point clouds, video, and multi-view scenes. Methods are commonly evaluated with Panoptic Quality on benchmarks such as COCO, Cityscapes, and ScanNet, while video and driving settings also test temporal consistency and robustness to adverse conditions. A central challenge is maintaining accurate, consistent instance identities while recognizing unfamiliar objects and meeting the efficiency and annotation constraints of real-world deployment.
Latest papers
38 papers7 papers in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.