Universal Image Segmentation with Mask2Former and OneFormer

The article introduces Mask2Former and OneFormer, two models designed for universal image segmentation tasks. Mask2Former employs a transformer-based architecture with a unified framework that integrates both instance and semantic segmentation, achieving state-of-the-art performance on benchmarks like COCO and ADE20K. OneFormer builds on this by introducing a single model capable of handling various segmentation tasks, which simplifies deployment for practitioners and enhances efficiency in training and inference across diverse datasets.