RoMa-Ω: What Feed-Forward 3D Models Know About Image Matching

David Nordström¹, Xinyue Zhang², Thibaut Loiseau³, Vincent Lepetit³, Fredrik Kahl¹

¹ Chalmers University of Technology
² Mobile Perception Lab, ShanghaiTech University
³ LIGM, École des Ponts, Univ. Gustave Eiffel, CNRS

We replace the DINOv3 backbone of RoMa v2 with features from the VGGT-Ω architecture. This small change leads to our model, RoMa-Ω, which achieves state-of-the-art performance on a wide range of benchmarks, surpassing the current best matchers RoMa and RoMa v2. In particular, \ours~surpasses RoMa on the difficult matching benchmarks WxBS and HardMatch, which RoMa v2 did not.

Overview

RoMa-Ω is a powerful but slow dense matcher that builds on RoMa v2. We change the DINOv3 backbone in RoMa v2 for VGGT-Ω and retrain. The resulting model is slightly stronger than RoMa v2 on most benchmarks and quite a bit stronger on difficult matching (like HardMatch and WxBS). As part of our paper, we also explore the zero-shot matching abilities of VGGT-Ω, which we release code for in this repo.

Updates

• [August 6, 2026] Initial public release of the code.

Setup/Install

Clone with submodules (this repo depends on VGGT-Ω via third_party/vggtomega):

git clone --recurse-submodules <this-repo-url>
# or, if already cloned:
git submodule update --init --recursive

Download the VGGT-Ω checkpoint from this link and place it at vggtomega.pt in the repo root (or pass --checkpoint explicitly).

In your python environment (tested on Linux python 3.12), run:

uv sync

Evaluation

We supply a script to evaluate on WxBS and HardMatch that automatically downloads the datasets. RUBIK is also supported but requires downloading the data in accordance with the repo.

python experiments/eval_sparse.py --name roma_omega --benchmark wxbs # mAA_10px: 0.7198
python experiments/eval_sparse.py --name roma_omega --benchmark hardmatch # mAA_10px: 0.5110

These results closely replicate those reported in the paper. You can also test the matching abilities of VGGT-Ω by using --name vggt_omega.

Dense PCK/EPE benchmarks:

python experiments/eval_dense.py

Note, the dense benchmarks require you to download the data yourself. For MegaDepth, follow the instructions in DKM.

Training

Unfortunately, for now you will have to download the datasets on your own. You can follow DKM for MegaDepth.

Both stages are run with torchrun on 8 GPUs (validated on A100s) and take roughly 3 days each on that setup:

torchrun --nproc_per_node=8 experiments/train_matcher.py --name my-run
torchrun --nproc_per_node=8 experiments/train_refiners.py --name my-run --matcher-run-path experiments/vggtroma/runs/my-run/<timestamp>

License

All our code is MIT license.

Acknowledgement

Code is based on RoMaV2 by Parskatt.

BibTeX

If you find our models useful, please consider citing our papers!

@inproceedings{nordstrom2026romaomega,
      title={RoMa-$\Omega$: What Feed-Forward 3D Models Know About Image Matching}, 
      author={David Nordström and Xinyue Zhang and Thibaut Loiseau and Vincent Lepetit and Fredrik Kahl},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV) Workshops},
      year={2026}
}