Mask2Former: Transformer-based universal image segmentation

Mask2Former is a computer vision model that performs image segmentation using transformer-based architecture with masked attention mechanisms. The system processes images through a backbone encoder and applies attention operations constrained by learned masks to generate segmentation outputs. It handles three segmentation task types (panoptic, instance, and semantic) through a single unified model architecture rather than task-specific variants. The codebase supports training and inference on major segmentation benchmarks including ADE20K, Cityscapes, COCO, and Mapillary Vistas, with additional support for video instance segmentation.