Why EoMT

• Encoder-only design, no pixel or transformer decoder
  
  
• Supports semantic, instance, and panoptic segmentation under a single framework
  
  
• Better throughput for real-world deployments
  
  
• Compatible with multiple DINOv3 backbones (vits16, vitb16, vitl16, etc.)

You can train this model on your own dataset with just a few lines of code:

import lightly_train

if __name__ == "__main__":
    # Train a model.
    lightly_train.train_semantic_segmentation(
        out="out/my_experiment",
 model="dinov2/vitl14-eomt",
 data={...},
    )

    # Load the trained model.
    model = lightly_train.load_model_from_checkpoint(
        "out/my_experiment/checkpoints/last.ckpt"
    )

    # Run inference.
    masks = model.predict("path/to/image.jpg")

‍

Benchmarks & Performance Across Datasets

Here’s a snapshot of how EoMT with DINOv3 stacks up in LightlyTrain across standard segmentation datasets.

Results reported on an NVIDIA T4 GPU. FPS values are given at the listed input resolutions.

In ADE20K, for example, dinov3/vitl16-eomt hits 59.1% mIoU, surpassing the DINOv2 EoMT baseline of 58.4%.

Why Use EoMT in LightlyTrain

• Simplicity + efficiency: No decoders or complex adapters. Fewer modules means faster inference and easier integration.
  
• Unified architecture: Supports semantic, instance, and panoptic segmentation.
  
• Performance + speed trade-off: With DINOv3 backbones, EoMT achieves competitive accuracy while maintaining strong throughput.
  
• Seamless pipeline support: Easily integrated into the LightlyTrain workflow, you can fine-tune your segmentation models with minimal boilerplate.

Try It Today

Ready to try it? Head over to the LightlyTrain docs to start experimenting:
👉 Semantic Segmentation — EoMT + DINOv3 docs

Dive deeper into how EoMT works and the architectural decisions behind it in our blog post: 👉 EoMT: Your ViT is Secretly an Image Segmentation Model