What it does

This repository collects and organizes resources for panoptic segmentation—the unified approach to labeling every pixel in an image as either “stuff” (background regions like sky or grass) or “things” (countable objects like cars or people). It gathers papers with links, official and unofficial code, dataset pointers, evaluation metrics, and benchmark leaderboards across COCO, Cityscapes, and Mapillary Vistas.

The interesting bit

The value is in the comparison tables. The maintainer tracks PQ (Panoptic Quality) scores and whether methods are end-to-end, letting you see at a glance which ResNet backbones actually move the needle and which papers are still missing code links.

Key highlights

• Covers 20+ papers from CVPR/ICCV/ECCV/AAAI 2018–2020, plus arXiv preprints
• Benchmark tables for three major datasets with consistent metric columns (PQ, SQ, RQ, mIoU, AP-Mask)
• Links to evaluation code (panopticapi, cityscapesScripts) and closed/open competitions
• Includes tutorials and a single blog link in Chinese

Caveats

• Coverage appears to stop around 2020; newer methods (EfficientPS was arXiv 2020) may not be fully updated
• Some benchmark cells are blank—unclear if metrics weren’t reported or weren’t collected
• No code in this repo itself; it’s purely a reference index

Verdict

Worth bookmarking if you’re entering panoptic segmentation research and need to map the landscape fast. Skip it if you want runnable implementations or state-of-the-art results from 2023 onward.