roma-outdoor

RoMa (outdoor) — weight mirror

Unmodified mirror of the two checkpoints that RoMa (CVPR 2024, Edstedt et al.) needs in order to run inference. Upstream ships no Hugging Face repo and downloads these via torch.hub at construction time; this repo exists so they can be fetched once with huggingface_hub and loaded offline afterwards.

Nothing here is retrained, converted, quantised or otherwise modified — the files are byte-identical to the upstream artefacts listed below.

Contents

Provenance

dinov2_vitl14_pretrain.pth is Meta's original DINOv2 ViT-L/14 state dict. It is not interchangeable with the transformers-format facebook/dinov2-large: RoMa ships its own DINOv2 implementation and expects this layout.

Usage

import torch
from huggingface_hub import snapshot_download
from romatch import roma_outdoor

local = snapshot_download("xboix/roma-outdoor")

# roma_outdoor() raises unless float32 matmul precision is "highest".
torch.set_float32_matmul_precision("highest")

# Passing both state dicts explicitly is what keeps this offline — leave either as None and
# romatch falls back to its torch.hub download.
model = roma_outdoor(
    device="cuda",
    weights=torch.load(f"{local}/roma_outdoor.pth", map_location="cpu"),
    dinov2_weights=torch.load(f"{local}/dinov2_vitl14_pretrain.pth", map_location="cpu"),
)
warp, certainty = model.match(im_a_path, im_b_path, device="cuda")
matches, certainty = model.sample(warp, certainty)

Licence

The mirrored weights keep their upstream licences, included verbatim:

• roma_outdoor.pth — MIT, © 2023 Johan Edstedt (LICENSE.RoMa).
• dinov2_vitl14_pretrain.pth — Apache-2.0, © Meta Platforms (LICENSE.DINOv2).

Citation

@inproceedings{edstedt2024roma,
  title={{RoMa: Robust Dense Feature Matching}},
  author={Edstedt, Johan and Sun, Qiyu and Bökman, Georg and Wadenbäck, Mårten and Felsberg, Michael},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024}
}