Panoptic FPN

A streamlined vision architecture from FAIR that unifies instance and semantic segmentation into a single, high-performance framework.

Panoptic FPN, engineered by the FAIR team (Kirillov, Girshick, He, and Dollár), delivers a unified approach to scene understanding by merging instance and semantic segmentation. The architecture utilizes a shared Feature Pyramid Network backbone to simultaneously identify things (individual objects like cars) and stuff (amorphous regions like roads or sky). By running on standard ResNet-50 or ResNet-101 backbones, it matches the performance of specialized models on the COCO and Cityscapes benchmarks without the complexity of dual pipelines. This efficiency makes it a robust choice for real-world applications requiring complete pixel-level context.

https://arxiv.org/abs/1901.02446