Add LoMa

[davnords]

Attempt to add the LoMa matcher from LoMa: Local Feature Matching Revisited.

Verification

When I benchmarked LoMa-B on InLoc, using the following simple script:

from pathlib import Path
from hloc import extract_features, match_features, localize_inloc

if __name__ == "__main__":
    print(f"Running matcher")
    dataset = Path("data/inloc")  # change this if your dataset is somewhere else

    pairs = Path("pairs/inloc")
    loc_pairs = pairs / "pairs-query-netvlad40.txt"  # top 40 retrieved by NetVLAD

    outputs = Path("outputs/inloc")  # where everything will be saved
    results = outputs / f"InLoc_loma_netvlad40.txt"  # the result file
    feature_conf = extract_features.confs["loma_inloc"]
    matcher_conf = match_features.confs["loma"]

    feature_path = extract_features.main(feature_conf, dataset, outputs)
    match_path = match_features.main(matcher_conf, loc_pairs, feature_conf["output"], outputs, overwrite=False)# True)
    localize_inloc.main(
        dataset, loc_pairs, feature_path, match_path, results, skip_matches=20
    )

I achieved the following results:

DUC1: 55.1 / 80.3 / 91.4 
DUC2: 72.5 / 87.8 / 88.5

This is similar to the results reported in the paper and an almost 20-point improvement on DUC2 (0.25m, 10°) compared to what is reported in the LightGlue paper.

Thank you very much for your work.

/ David

[davnords]

@sarlinpe Perhaps this matcher be interest you :)

LoMa.

/ David

[omaralvarez]

This matcher would be a great addition, hoping it gets merged!

[Parskatt]

@sarlinpe anything else required for merge?

[omaralvarez]

I have been testing the implementation and I found an issue with the Dockerfile of hloc. For up to date colmap, hloc, and LoMa, the Python version needs to be bumped to 3.10, if not the docker build errors out. Additionally, probably due to being based in the colmap image it is missing git.

In python 3.10, dataclasses is already included in the Python standard library, leading to a dependency conflict when installing lomatch @ git+https://github.com/davnords/LoMa.

Just removing the dataclasses in the .toml in LoMa would fix this issue. I could create another pull request with the updated Dockerfile.

[Parskatt]

Now fixed.

[omaralvarez]

I confirm now docker works well with the change and my Dockerfile pull request #497 . I have also tested LoMa with a difficult reconstruction and it works really well.

[omaralvarez]

Shouldn't we make available through config all LoMa weights (loma_R, loma_L, etc.)? It would be great to be able to use the LoMa-R version. I think right now LoMa-R is missing, please correct me if I am wrong.

[davnords]

Thank you @omaralvarez for testing. I am glad LoMa performed well on difficult reconstructions. I have now added LoMa-R to the configs.