Practical overview
When it fits
Use EoMT when you want a code-based workflow for experimenting with Vision Transformer image segmentation, rather than a hosted inference service.
Who it is for
- Computer-vision researchers exploring transformer-based image segmentation.
- Practitioners who need code for training or evaluating segmentation models on documented datasets.
- Developers adapting a pre-trained EoMT checkpoint to a dataset with different classes.
Documented uses
- Train an EoMT model for COCO panoptic segmentation.
- Evaluate a pre-trained checkpoint against a dataset.
- Fine-tune a checkpoint while skipping its classification head for a different class set.
- Run quick inference and visualization through the provided notebook.
Requirements & limits
The public GitHub repository lists an MIT license.
- The documented environment uses Conda and Python 3.13.2, followed by installation from requirements.txt.
- The example COCO panoptic training and evaluation commands use four GPUs and a total batch size of 16.
- The README says dataset archives can remain zipped and be supplied by directory through --data.path.
- The inference notebook is described as automatically downloading pre-trained models.
- DINOv3 configurations expect delta weights by default; the README documents an option to use absolute weights.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up EoMT for segmentation experiments
Create the documented Conda environment and install the project dependencies.
Before you start
- A shell with Conda installed; the README recommends Miniconda if needed.
- The EoMT repository and its requirements.txt file.
Steps
Create a Conda environment with Python 3.13.2 using the documented environment name eomt.
Activate the eomt environment.
Install dependencies with python3 -m pip install -r requirements.txt.
Expected outcome
The documented Python environment and project dependencies are installed.
Guide limitations
- These steps cover setup only; training and evaluation require additional datasets, configurations and model weights.
- Weights & Biases login is documented for enabling experiment logging, not as a prerequisite for environment setup.
Prepare COCO archives for EoMT
Download the COCO 2017 image and annotation archives and place them for use through the project data path.
Before you start
- A shell with wget available.
- A directory to hold the downloaded dataset archives.
Steps
Download train2017.zip and val2017.zip from the COCO images location shown in the README.
Download annotations_trainval2017.zip and panoptic_annotations_trainval2017.zip from the COCO annotations location.
Keep the downloaded files zipped, place them in the chosen directory, and provide that directory through the --data.path argument.
Expected outcome
The COCO archives are prepared in the layout expected by the documented data-path workflow.
Guide limitations
- This recipe prepares data only; it does not run training or evaluation.
- The repository gives separate dataset preparation instructions for ADE20K and Cityscapes.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- tue-mps/eomt · Official homepage title · retained in its source language