Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
4
31
outdoor
object
683.49 335.11 571.26 329.29
321.34 338.29 385.38 353.25
205.96 325.54 330.90 353.25
878.99 277.70 704.27 260.60
917.45 300.02 725.10 276.57
369.41 236.24 366.15 310.12
125.84 298.43 295.65 340.47
508.83 371.78 459.09 374.01
864.57 335.11 683.43 310.12
465.56 437.16 500.76 378.81
683.49 335.11 571.26 329.29
321.34 338.29 385.38 353.25
205.96 325.54 330.90 353.25
878.99 277.70 704.27 260.60
917.45 300.02 725.10 276.57
369.41 236.24 366.15 310.12
125.84 298.43 295.65 340.47
508.83 371.78 459.09 374.01
864.57 335.11 683.43 310.12
465.56 437.16 500.76 378.81
outdoor
object
doppelgänger
547.30 2167.30 3214.34 1316.46
4722.91 821.54 653.74 552.24
3917.14 1467.51 1168.58 928.53
670.11 2053.66 3146.34 1250.51
4150.78 1282.09 1032.58 835.43
4689.96 971.07 671.22 637.58
4713.92 1844.32 663.45 1153.53
4606.09 1312.00 721.73 823.79
4366.45 1000.98 888.81 668.62
100.69 398.89 911.77 373.55
868.89 151.20 185.44 156.70
720.65 270.09 331.48 263.48
123.28 377.97 892.48 354.84
763.64 235.97 292.90 237.06
862.83 178.72 190.40 180.92
867.24 339.44 188.19 327.32
847.40 241.47 204.72 233.76
803.31 184.23 252.12 189.72
outdoor
time
object
107.72 187.18 89.77 178.20
438.07 280.59 395.11 265.64
515.68 228.92 469.70 221.92
542.55 210.03 492.58 205.03
537.57 198.11 489.59 194.10
453.00 292.52 408.04 278.55
171.40 103.70 147.46 131.50
381.35 280.59 343.39 267.62
140.56 194.13 122.59 185.16
244.04 280.59 211.11 278.55
209.22 268.67 181.27 263.65
107.72 187.18 89.77 178.20
438.07 280.59 395.11 265.64
515.68 228.92 469.70 221.92
542.55 210.03 492.58 205.03
537.57 198.11 489.59 194.10
453.00 292.52 408.04 278.55
171.40 103.70 147.46 131.50
381.35 280.59 343.39 267.62
140.56 194.13 122.59 185.16
244.04 280.59 211.11 278.55
209.22 268.67 181.27 263.65
outdoor
time
object
655.81 620.44 197.13 122.36
692.21 620.44 214.04 121.37
697.96 643.40 217.02 132.28
743.94 639.57 237.91 130.30
533.19 700.78 142.43 167.99
876.13 712.26 304.55 164.02
734.36 704.61 240.89 166.01
799.50 607.05 259.79 113.43
868.47 542.02 301.56 76.73
1075.38 545.84 414.95 71.77
1299.54 561.15 532.31 75.74
1207.58 568.80 477.61 79.71
897.21 712.26 312.50 164.02
646.23 645.31 194.15 136.25
377.27 357.05 197.13 122.36
398.22 357.05 214.04 121.37
401.52 370.26 217.02 132.28
427.97 368.06 237.91 130.30
306.73 403.28 142.43 167.99
504.02 409.89 304.55 164.02
422.46 405.49 240.89 166.01
459.94 349.34 259.79 113.43
499.61 311.92 301.56 76.73
618.65 314.12 414.95 71.77
747.60 322.93 532.31 75.74
694.70 327.33 477.61 79.71
516.15 409.89 312.50 164.02
371.76 371.36 194.15 136.25
outdoor
End of preview. Expand in Data Studio

HardMatch (unofficial mirror)

This repository is an unofficial mirror. It redistributes the HardMatch benchmark exactly as published by its authors (David Nordström, Johan Edstedt, and co-authors). It is not affiliated with, endorsed by, or maintained by the HardMatch/LoMa authors. The authoritative source is the official GitHub repository: https://github.com/davnords/HardMatch. If you are an author or a rights holder and would like content changed or removed, please open a discussion on this repository or an issue on the official GitHub repository, and it will be handled promptly.

Dataset summary

HardMatch is an extremely difficult image matching benchmark with 1,000 hand-annotated image pairs, organized into 100 object/scene categories and tagged with 15 scene-attribute labels. It was released as part of the LoMa paper (ECCV 2026, Oral).

Image pairs 1,000
Categories 100 (10 pairs each)
Scene-attribute labels aerial, building, celestial, doppelgänger, drawing, dynamic, illumination, indoor, nature, object, outdoor, season, texture, time, viewpoint
Images 2,000 JPEGs, longer side 1024 px
Size ~685 MB
Sources Images scraped from Wikimedia Commons
Image licenses Mixed permissive Commons licenses — see each pair's metadata.json
Annotation license Released by the HardMatch/LoMa authors; code is MIT

Repository layout

Each pair is stored in its own directory:

<category>/<object>/pair_XXXXX/
├── A_lr.jpg           # low-resolution image A (longer side 1024 px)
├── B_lr.jpg           # low-resolution image B
├── A_metadata.json    # Wikimedia Commons metadata for A (author, license, URL, ...)
├── B_metadata.json    # Wikimedia Commons metadata for B
├── categories.txt     # scene-attribute labels, one per line
├── corresp.txt        # ground-truth correspondences in ORIGINAL image coordinates
└── corresp_lr.txt     # ground-truth correspondences in the distributed (1024 px) images

corresp*.txt contains one correspondence per line with four space-separated floats: x_A y_A x_B y_B (pixel coordinates, 0-indexed top-left origin). Use corresp_lr.txt together with A_lr.jpg / B_lr.jpg; corresp.txt is provided for reference only because the full-resolution originals are not redistributed.

All 100 categories

airplane, amphitheater, amphora, anchor, aquatic, aqueduct, atrium, bell_tower, bridge, cannon, canyon, cargo_ship, cassiopeia, castle, cathedral, cave, cave_interior, chapel, château, chimney, church, citadel, city_gate, clock_tower, corridor, crane, dam, doll, drone, earth, famous_animals, fjord, football, fortress, fountain, garden, geyser, glacier, gymnastics, harbor, hot_spring, island, isolated_monastery, library_interior, lighthouse, marina, mausoleum, moon, moons_of_jupiter, mosque, obelisk, oil_refinery, old_town, opera_houses, orchard, orion, pagoda, palace, park, pedestrian_street, pier, planetarium, power_lines, power_plant, pyramid, railway_station, relief_sculpture, ruins, rune_stones, sailboat, sarcophagus, satellite_dish, sculpture, sea_cliff, skyscraper, stairwell, statue, summit, suspension_bridge, synagogue, tank, telephone_booth, temple, tennis, tower, town_square, trees, triumphal_arch, tunnel, ursa_major, vase, volcano, water_tower, water_well, waterfall, waterfront, weapon, wind_farm, windmill, zeppelin

How to download and use

Download the full dataset:

from huggingface_hub import snapshot_download

root = snapshot_download(repo_id="Realcat/hardmatch", repo_type="dataset")
# root/airplane/.../pair_00010/A_lr.jpg, etc.

Use it with the official benchmark code:

pip install hardmatch
from hardmatch.dataset import HardMatch

dataset = HardMatch(HardMatch.Cfg(data_root=root, download=False))

This repository uses the original directory layout, so it is a drop-in replacement for the zip file downloaded automatically by hardmatch (the only difference is that the data is already extracted). It is intentionally not converted to Parquet: the ground-truth correspondences are per-pair free-form text files, and preserving the original file paths keeps existing tooling compatible.

To use only some pairs, pass allow_patterns / ignore_patterns to snapshot_download, e.g. allow_patterns=["cathedral/**"] (categories are top-level directories).

Licensing, attribution, and compliance

You are responsible for complying with the license of each image you use or redistribute. The images were scraped from Wikimedia Commons and retain their original licenses; this mirror grants no additional rights.

Every pair ships its own license information. Check the metadata.json files before using or redistributing an image:

import json

meta = json.load(open(f"{pair_dir}/A_metadata.json"))
print(meta["license"], meta["license_url"])
print(meta["author"], meta["original_url"])
print(meta["attribution_required"])

Relevant fields: license, license_url, author, attribution_required, original_url, file_title, and the full extmetadata_raw block returned by the Commons API. Across the 2,000 images:

  • Roughly 84% (1,672) have attribution_required == "true".
  • The most common licenses are CC BY-SA 4.0 (755), CC BY-SA 3.0 (291), Public domain (201), CC BY-SA 2.0 (181), CC BY 2.0 (158), CC0 (111), CC BY 4.0 (105), and CC BY 3.0 (98), plus ~25 other Commons licenses.
  • 6 images have no attribution_required value and 1 has no license field; inspect their metadata.json / original_url and treat them as requiring attribution unless the source page says otherwise.

If you redistribute or publish results based on these images:

  1. Attribution-required licenses (CC BY, CC BY-SA, and similar): credit the author, name the license, link to the license, and indicate if changes were made. author may contain HTML from Commons; strip the tags for a plain-text credit.
  2. Share-alike licenses (CC BY-SA): derivative works and redistributed datasets containing these images must be released under the same license.
  3. Public domain / CC0 / "No restrictions": no conditions, but crediting the source is good practice.
  4. Do not remove or alter the A_metadata.json / B_metadata.json files when redistributing the data; they carry the attribution information.
  5. Cite the paper (below) when using the benchmark — this is a scientific attribution requirement, independent of the image licenses.

Known issues in this snapshot

The files are uploaded without modification. This snapshot contains two anomalies:

  • bell_tower/Zvonnitsa_in_Prokhorovka_Battlefield/pair_0001/ contains metadata and correspondences but no images or corresp_lr.txt (likely removed upstream for licensing reasons). The official loader skips it, yielding 999 loadable pairs.
  • tower/Q118464737/pair_00692/ is missing categories.txt; all other files are present.
  • The official archive also contains two empty category directories, gargoyle/ and wild_animals/, which are not tracked by Git and therefore do not appear here.

Citation

@inproceedings{nordstrom2026loma,
      title={LoMa: Local Feature Matching Revisited},
      author={David Nordström and Johan Edstedt and Georg Bökman and Jonathan Astermark and Anders Heyden and Viktor Larsson and Mårten Wadenbäck and Michael Felsberg and Fredrik Kahl},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
      year={2026}
}

Links

Downloads last month
60

Paper for Realcat/hardmatch