Johan Edstedt · Qiyu Sun · Georg Bökman · MÄrten WadenbÀck · Michael Felsberg
RoMa is the robust dense feature matcher capable of estimating pixel-dense warps and reliable certainties for almost any image pair.
In your python environment (tested on Linux python 3.10), run:
pip install -e .
We provide two demos in the demos folder. Here's the gist of it:
from romatch import roma_outdoor
roma_model = roma_outdoor(device=device)
# Match
warp, certainty = roma_model.match(imA_path, imB_path, device=device)
# Sample matches for estimation
matches, certainty = roma_model.sample(warp, certainty)
# Convert to pixel coordinates (RoMa produces matches in [-1,1]x[-1,1])
kptsA, kptsB = roma_model.to_pixel_coordinates(matches, H_A, W_A, H_B, W_B)
# Find a fundamental matrix (or anything else of interest)
F, mask = cv2.findFundamentalMat(
kptsA.cpu().numpy(), kptsB.cpu().numpy(), ransacReprojThreshold=0.2, method=cv2.USAC_MAGSAC, confidence=0.999999, maxIters=10000
)
New: You can also match arbitrary keypoints with RoMa. See match_keypoints in RegressionMatcher.
By default RoMa uses an initial resolution of (560,560) which is then upsampled to (864,864). You can change this at construction (see roma_outdoor kwargs). You can also change this later, by changing the roma_model.w_resized, roma_model.h_resized, and roma_model.upsample_res.
roma_model.sample_thresh controls the thresholding used when sampling matches for estimation. In certain cases a lower or higher threshold may improve results.
The experiments in the paper are provided in the experiments folder.
- First follow the instructions provided here: https://github.com/Parskatt/DKM for downloading and preprocessing datasets.
- Run the relevant experiment, e.g.,
torchrun --nproc_per_node=4 --nnodes=1 --rdzv_backend=c10d experiments/roma_outdoor.py
python experiments/roma_outdoor.py --only_test --benchmark mega-1500
All our code except DINOv2 is MIT license. DINOv2 has an Apache 2 license DINOv2.
Our codebase builds on the code in DKM.
If you find that RoMa is too heavy, you might want to try Tiny RoMa which is built on top of XFeat.
from romatch import tiny_roma_v1_outdoor
tiny_roma_model = tiny_roma_v1_outdoor(device=device)
Mega1500:
AUC@5 | AUC@10 | AUC@20 | |
---|---|---|---|
XFeat | 46.4 | 58.9 | 69.2 |
XFeat* | 51.9 | 67.2 | 78.9 |
Tiny RoMa v1 | 56.4 | 69.5 | 79.5 |
RoMa | - | - | - |
Mega-8-Scenes (See DKM):
AUC@5 | AUC@10 | AUC@20 | |
---|---|---|---|
XFeat | - | - | - |
XFeat* | 50.1 | 64.4 | 75.2 |
Tiny RoMa v1 | 57.7 | 70.5 | 79.6 |
RoMa | - | - | - |
IMC22 :'):
mAA@10 | |
---|---|
XFeat | 42.1 |
XFeat* | - |
Tiny RoMa v1 | 42.2 |
RoMa | - |
If you find our models useful, please consider citing our paper!
@article{edstedt2024roma,
title={{RoMa: Robust Dense Feature Matching}},
author={Edstedt, Johan and Sun, Qiyu and Bökman, Georg and WadenbÀck, MÄrten and Felsberg, Michael},
journal={IEEE Conference on Computer Vision and Pattern Recognition},
year={2024}
}