EoMT
tue-mpsEoMT (Encoder-only Mask Transformer) is an image segmentation model that turns a plain Vision Transformer into a segmenter without adapters or decoders, by passing segmentation queries through the encoder alongside the image patches. Built on DINOv2 or DINOv3 backbones, it handles semantic, instance and panoptic segmentation and is available in Hugging Face Transformers. It comes from the Mobile Perception Systems Lab at Eindhoven University of Technology.
Openness
4 medium confidence- weights
- open (DINOv2- and DINOv3-based COCO, ADE20K and Cityscapes checkpoints on the Hub, ungated; the DINOv3 ones are deltas over Meta's DINOv3 weights)
- data
- documented-not-released (fine-tuned on COCO, ADE20K and Cityscapes from DINOv2 and DINOv3 backbones pretrained on Meta's unreleased LVD-142M and LVD-1689M corpora)
- code
- open (training and evaluation code and configs)
- license
- MIT (code and checkpoints)+DINOv3-License (the DINOv3 backbone weights the DINOv3-based checkpoints are applied to)
EoMT's code and checkpoints are MIT, and the training code and configs for COCO, ADE20K and Cityscapes are public. The segmenters are fine-tuned from Meta's DINOv2 and DINOv3 backbones, whose pretraining images Meta has not released, and the DINOv3-based checkpoints are published as differences that need Meta's DINOv3 weights, which come under Meta's own license.
- https://huggingface.co/api/models?author=tue-mps&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
tue-mps EoMT checkpoints for COCO, ADE20K and Cityscapes, DINOv2- and DINOv3-based, all gated false and tagged mit.
- https://raw.githubusercontent.com/facebookresearch/dinov2/main/MODEL_CARD.md recorded 2026-11-12
DINOv2 model card: "Training data: LVD-142M (see paper)"; no release of the dataset.
- https://raw.githubusercontent.com/facebookresearch/dinov3/main/MODEL_CARD.md recorded 2026-11-12
DINOv3 model card: "Web dataset (LVD-1689M): a curated dataset of 1,689 millions of images extracted from a large data pool of 17 billions web images"; no release of the dataset.
- https://raw.githubusercontent.com/tue-mps/eomt/master/LICENSE recorded 2026-11-12
LICENSE is the MIT License, "Copyright (c) 2025 Mobile Perception Systems Lab at TU/e".
- https://raw.githubusercontent.com/tue-mps/eomt/master/model_zoo/dinov3.md recorded 2026-11-12
DINOv3 model zoo: EoMT-S, B and L COCO panoptic, COCO instance and ADE20K semantic configs with pytorch_model.bin downloads from tue-mps/*_dinov3 repositories.
- https://raw.githubusercontent.com/tue-mps/eomt/master/README.md recorded 2026-11-12
README: COCO, ADE20K and Cityscapes download steps; "To train EoMT from scratch, run: python3 main.py fit"; "When using DINOv3-based configurations, the code expects delta weights relative to DINOv3 weights by default"; pre-trained weights for DINOv2- and DINOv3-based models.
- https://ungh.cc/repos/tue-mps/eomt/files/master recorded 2026-11-12
Repository tree of 75 paths: configs for dinov2 and dinov3 on COCO, ADE20K and Cityscapes, and a training/ package with the Lightning module and mask-classification losses.
Adoption
2 medium confidenceHugging Face downloads over the trailing 30 days for the four most-used EoMT checkpoints, led by the COCO panoptic large model. Weights fetched through the repository's own download links count the same way, since they are served from the Hub.
- https://huggingface.co/api/models?author=tue-mps&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
coco_panoptic_eomt_large_640 42,322, eomt-dinov3-ade-semantic-large-512 21,494, ade20k_semantic_eomt_large_512 5,231 and coco_instance_eomt_large_1280 4,762 downloads in the trailing 30 days (73,809 together).
Capability
3 high confidenceEach EoMT checkpoint labels pixels with the classes of the benchmark it was trained on, so a new set of classes needs a training run, as with RT-DETR's COCO detectors.
- https://raw.githubusercontent.com/tue-mps/eomt/master/README.md recorded 2026-11-12
README: DINOv3-based EoMT "Panoptic Segmentation: Up to 58.9 PQ on COCO", "Instance Segmentation: Up to 49.9 mAP on COCO", "Semantic Segmentation: Up to 59.5 mIoU on ADE20K"; fine-tuning with --model.load_ckpt_class_head False on a dataset with different classes.
Verified 2026-11-12