Image Matching Models

A unified API for quickly and easily trying 23 (and growing!) image matching models.

Example Results

Compare matching models across various scenes. For example, we show SIFT-LightGlue and LoFTR matches on pairs:

(1) outdoor, (2) indoor, (3) satellite remote sensing, (4) paintings, and (5) a false positive.

SIFT-LightGlue

LoFTR

Install

To install this repo run

git clone --recursive https://github.com/gmberton/image-matching-models

You can install this package for use in other scripts/notebooks with the following

cd image-matching-models
python -m pip install -e .

Use

You can use any of the matchers with

from matching import get_matcher

device = 'cuda' # 'cpu'
matcher = get_matcher('superglue', device=device)  # Choose any of our ~20 matchers listed below
img_size = 512

img0 = matcher.image_loader('path/to/img0.png', resize=img_size)
img1 = matcher.image_loader('path/to/img1.png', resize=img_size)

result = matcher(img0, img1)
num_inliers, H, mkpts0, mkpts1 = result['num_inliers'], result['H'], result['mkpts0'], result['mkpts1']
# result.keys() = ['num_inliers', 'H', 'mkpts0', 'mkpts1', 'inliers0', 'inliers1', 'kpts0', 'kpts1', 'desc0', 'desc1']

You can also run this as a standalone script, which will perform inference on the the examples inside ./assets. It is possible to specify also resolution and num_keypoints. This will take a few seconds also on a laptop's CPU, and will produce the same images that you see above.

python main.py --matcher sift-lg --device cpu --log_dir output_sift-lg

Where sift-lg will use SIFT + LightGlue.

You can choose any of the following methods: loftr, [sift, superpoint, disk, aliked, dedode, doghardnet, gim]-lg, roma, dedode, steerers, [sift, orb, doghardnet]-nn, patch2pix, patch2pix_superglue, superglue, r2d2, d2net, duster, gim-dkm, xfeat, omniglue

The script will generate an image with the matching keypoints for each pair, under ./output_sift-lg.

All the matchers can run on GPU, and most of them can run both on GPU or CPU. A few can't run on CPU.

Use on your own images

To use on your images you have three options:

1. create a directory with sub-directories, with two images per sub-directory, just like ./assets/example_pairs. Then use as python main.py --input path/to/dir
2. create a file with pairs of paths, separate by a space, just like assets/example_pairs_paths.txt. Then use as python main.py --input path/to/file.txt
3. import the matcher package into a script/notebook and use from there, as in the example above

Models

Model	Code	Paper	GPU Runtime (s/img)	CPU Runtime (s/img)
OmniGlue (CVPR '24)	Official	arxiv	❌	6.351
xFeat (CVPR '24)	Official	arxiv	0.027	0.048
GIM (ICLR '24)	Official	arxiv	0.077 (+LG) / 1.627 (+DKMv3)	5.321 (+LG) / 20.301 (+DKMv3)
RoMa (CVPR '24)	Official	arxiv	0.453	18.950
Dust3r (CVPR '24)	Official	arxiv	3.639	26.813
DeDoDe (3DV '24)	Official	arxiv	0.311 (+MNN)/ 0.218+(LG)	❌
Steerers (arxiv '24)	Official	arxiv	0.150	❌
LightGlue* (ICCV '23)	Official	arxiv	0.417 / 0.093 / 0.184 / 0.128	2.828 / 8.852 / 8.100 / 8.128
SiLK (ICCV '23)	Official	arxiv	0.694	3.733
LoFTR (CVPR '21)	Official / Kornia	arxiv	0.722	2.36
Patch2Pix (CVPR '21)	Official / IMT	arxiv	--	--
SuperGlue (CVPR '20)	Official / IMT	arxiv	0.0894	2.178
R2D2 (NeurIPS '19)	Official / IMT	arxiv	0.429	6.79
D2Net (CVPR '19)	Official / IMT	arxiv	0.600	1.324
SIFT- NN (IJCV '04)	OpenCV	pdf	0.124	0.117
ORB- NN (ICCV '11)	OpenCV	ResearchGate	0.088	0.092
DoGHardNet (NeurIPS '17)	IMT / Kornia	arxiv	2.697 (+NN) / 0.526 +(LG)	2.438(+NN) / 4.528 (+LG)

Our implementation of Patch2Pix (+ Patch2PixSuperGlue), R2D2, and D2Net are based on the Image Matching Toolbox (IMT). LoFTR and DeDoDe-Lightglue are from Kornia. Other models are based on the offical repos above.

Runtime benchmark is the average of 5 iterations over the 5 pairs of examples in the assets/example_pairs folder at image size 512x512. Benchmark is done using benchmark.py on an NVIDIA RTX A4000 GPU. Results rounded to the hundredths place.

* LightGlue model runtimes are listed in the order: SIFT, SuperPoint, Disk, ALIKED

Adding a new method

To add a new method simply add it to ./matching. If the method requires external modules, you can add them to ./third_party with git submodule add: for example, I've used this command to add the LightGlue module which is automatically downloaded when using --recursive

git submodule add https://github.com/cvg/LightGlue third_party/LightGlue

This command automatically modifies .gitmodules (and modifying it manually doesn't work).

Note

This repo is not optimized for speed, but for usability. The idea is to use this repo to find the matcher that best suits your needs, and then use the original code to get the best out of it.

Acknowledgements

Special thanks to the authors of the respective works that are included in this repo (see their papers above). Additional thanks to @GrumpyZhou for developing and maintaining the Image Matching Toolbox, which we have wrapped in this repo, and the maintainers of Kornia.

Cite

This repo was created as part of the EarthMatch paper.

@InProceedings{Berton_2024_EarthMatch,
    author    = {Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, Carlo Masone},
    title     = {EarthMatch: Iterative Coregistration for Fine-grained Localization of Astronaut Photography},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2024},
}