eomt

Why we included this project

EoMT is a CVPR 2025 Highlight paper that questions how much architecture image segmentation really needs. It takes a plain, pre-trained Vision Transformer and uses it to encode both image patches and segmentation queries as tokens, with no adapters and no decoder. That simplicity still lands accuracy close to state-of-the-art methods built on task-specific components, while running up to 4x faster with a ViT-L backbone. The repo ships official training and evaluation code for panoptic, instance, and semantic segmentation on COCO, ADE20K, and Cityscapes, along with configs for DINOv2 and DINOv3 backbones and a model zoo of pretrained weights. Teams can benchmark it as a lean baseline against their existing pipeline, and the companion VidEoMT work extends the same minimal design to video segmentation.