turhancan97/wroclaw-dwarves

Wrocław Dwarves — Fine-Grained Instance Retrieval

Wrocław is scattered with several hundred small bronze dwarf statues — krasnale — installed across the city since 2001. They are individually sculpted, but they share a visual vocabulary: the same scale, the same material, the same crouching poses and hand-held props. Telling one from another is therefore a fine-grained instance recognition problem rather than a classification one. Every statue here belongs to the same semantic category; what distinguishes them is a hat, a tool, a pose.

This is a retrieval benchmark: 1,691 reference photographs of 306 distinct statues from Wikimedia Commons, with per-image attribution and licence metadata, per-statue coordinates where they could be established, precomputed embeddings, and a hash-pinned leave-one-out protocol. It is deliberately not a training set. The median class has four photographs and the largest has thirty-one — enough to retrieve against, not enough to train on. The point of the benchmark is that with a good enough general-purpose embedding you do not need to train at all.

The question it was built to answer is how identification accuracy decays as the candidate pool grows, and whether narrowing that pool by location — as a phone app would — actually helps. It does not: proximity-based pools are consistently harder than random pools of the same size, because sculptors install related pieces near each other, so a small radius selects for a statue's own lookalikes. The full method and every result artifact are in the source repository.

Task and protocol

Given a query photograph of one statue, rank the 306 known statues by how likely each is to be the one depicted. Candidates are ranked as distinct statues, each represented by its best-matching reference photograph — the list an identification tool would actually show.

Leave-one-out. Each of the 1,691 photographs is used once as a query and scored against the other 1,690, with the query withheld from its own class. The 1,691 folds are enumerated explicitly in the leave_one_out config so results are comparable across methods rather than dependent on a re-derived split. Intervals are 95% Wilson score intervals over the queries.

Baselines

Off-the-shelf embeddings identify these statues well with no fine-tuning:

	DINOv2 (facebook/dinov2-base)	CLIP (openai/clip-vit-base-patch32)
top-1	93.1% [91.8, 94.3]	82.9% [81.0, 84.6]
top-5	95.7% [94.6, 96.6]	90.7% [89.2, 92.0]
MRR	0.943	0.866

Both backbones are frozen and pinned to immutable revisions; images are embedded and ranked by cosine similarity. No fine-tuning, no probe, no metric learning. The ten-point gap is the expected ordering — DINOv2's self-supervised objective preserves instance-level detail, while CLIP's language alignment pulls toward semantic categories, and every statue here is the same semantic category.

Adding geometric verification on top — RANSAC inlier counts between the query and its top-10 candidates, blended into the similarity — lifts DINOv2 to 94.0% and CLIP to 86.3%, gaining 0.9 and 3.4 points. Geometry is worth four times as much to the weaker backbone, for the same reason the limitations below give.

Four secondary results are worth knowing before using this data. A per-fold linear probe is worth nothing to DINOv2 and gains CLIP about three points — where retrieval is already strong, training is not the right tool. Proximity-based candidate pools lose to random pools of the same size at every radius measured. Thresholding similarity for open-set rejection does not survive the larger pool: it worked at 23 classes and does not at 306, which retracts a positive result an earlier version of this project published. And identifying a statue from somebody else's photograph is much harder than the headline suggests — see the limitations.

Configurations

Config	Split	Rows	What it is
default	reference	1691	Photographs, metadata and rights columns
metadata	reference	1691	The same rows without the pixels
classes	reference	306	One row per statue, with coordinates
leave_one_out	test	1691	The evaluation folds
embeddings_dinov2	reference	1691	Precomputed dinov2 vectors
embeddings_clip	reference	1691	Precomputed clip vectors

from datasets import load_dataset

images = load_dataset("turhancan97/wroclaw-dwarves", "default", split="reference")
vectors = load_dataset("turhancan97/wroclaw-dwarves", "embeddings_dinov2", split="reference")
folds = load_dataset("turhancan97/wroclaw-dwarves", "leave_one_out", split="test")

The split is called reference, not train, because nothing here is trained. Every config is sorted by the same key, so row i of default and row i of either embeddings config are the same photograph.

Only the config you ask for is downloaded, so working from vectors alone costs a few megabytes rather than the full corpus.

Backbone	Model	Revision	Dimensions
dinov2	facebook/dinov2-base	f9e44c814b77	768
clip	openai/clip-vit-base-patch32	c237dc49a33f	512

Licensing and attribution

Every photograph keeps the licence it arrived with. See LICENSES.md for the full inventory and ATTRIBUTION.md for the credits; credits.csv is the same ledger in machine-readable form, downloadable without a parquet reader.

The obligations travel in the data. Each row carries author, license, license_spdx, license_url, source_url, commons_file, modified and a ready-to-paste attribution_text. 1,538 of the 1,691 files are modified — downscaled to at most 2000 px on the long edge and re-encoded; EXIF not preserved — and 153 are byte-identical to the Commons original; the modified column is derived by comparing the stored bytes against commons_sha1, so it can be checked rather than trusted.

Filtering by licence is one line, if you need a permissive subset:

permissive = images.filter(lambda row: row["license_spdx"] in {"CC0-1.0", "CC-BY-4.0"})

No further terms are imposed here, and none may be added by anyone redistributing it.

Personal and sensitive information

The dataset documents public sculpture, not people, and no photograph was selected for the people in it. But these are photographs taken in city streets, so passers-by may appear incidentally and may be identifiable. No face detection or blurring was applied, and the frequency was not measured. Every file was already publicly hosted on Wikimedia Commons before inclusion here, and each carries the Commons page it came from.

The EXIF of the 1,538 re-encoded files was not preserved, so the photographers' camera serial numbers and GPS traces are absent from them; the 153 untouched files retain whatever EXIF Commons holds. Per-photograph coordinates in the metadata are camera positions recorded by Commons, already public there, and are used to locate statues rather than photographers.

If you appear in one of these photographs and want it removed, the removal path below applies to you as much as to a rights-holder, and no legal argument is required.

Depicted works and freedom of panorama

The statues are contemporary sculptures by living artists, protected by copyright independently of the photographs. Wikimedia Commons hosts photographs of them under Poland's freedom-of-panorama provision (Art. 33(1) of the Polish Copyright Act), which permits publishing images of works permanently displayed in public places. Poland's provision is broad; the United States has no equivalent exemption for sculpture, and this platform is US-hosted.

The Creative Commons licences recorded per row cover the photographs, granted by the photographers. They do not, and cannot, grant any rights in the sculptures depicted. A reuse that would require the sculptor's permission in your jurisdiction still requires it here. This is a factual disclosure rather than legal advice, and it is recorded because it is the one rights question the per-file licences do not answer.

Removal requests. If you are a sculptor, photographer, rights-holder or their representative and want a photograph removed, open a discussion on this repository. Requests are honoured without requiring a legal argument. Removal is implemented upstream rather than patched here: the file is added to the source project's tracked data/image-review.json exclusions, the manifest, folds, embeddings and credit files are regenerated, and a new revision is pushed.

Limitations

• The reference photographs are not phone photographs. They are Commons uploads: mostly good light, considered framing, taken by someone who meant to document the statue. Using the photographs that were themselves shot on phones as a proxy, DINOv2 loses 5.3 top-1 points and CLIP 15.6 — and that is a lower bound, since those are still Commons uploads. Treat 93.1% as a ceiling a phone-camera application will not reach.
• The corpus is concentrated in few hands, and it costs a measurable amount. 122 photographers contributed, but Pnapora took 715 photographs and Fallaner took 424 — together 67.4% of the dataset. Near-duplicates from a single visit were never removed, so a method can score partly by recognising a camera and a processing style rather than a statue. Withholding each query's own photographer and comparing against a size-matched random control puts the part attributable to the photographer at 2.6 top-1 points for DINOv2 and 13.2 for CLIP — so the headline is inflated, modestly for DINOv2 and substantially for CLIP. Two further caveats follow from the same concentration: 125 of the 306 classes have a single photographer, so a third of the dataset cannot be evaluated cross-photographer at all, and cross-photographer accuracy is far below the headline — 81.8% for DINOv2 and 54.1% for CLIP. If your use case means matching a photograph against references somebody else took, those are the numbers to plan around.
• Most coordinates are derived, not stated. 271 of the 294 placed statues are located from where photographers stood rather than from a Wikidata P625 statement; validated against the classes that have both, the median error is 9 m. 12 statues are unplaced. coordinate_source says which is which. Do not present these as official locations.
• Two classes overlap with a group class. Grajek i Meloman is a class, and so are Grajek and Meloman; likewise Ogrodnik i Kierownik and Ogrodnik. Their photographs are not byte-identical so the duplicate guard does not catch them, and they duly appear among the confusions.
• Class sizes are uneven — three to thirty-one photographs, median four. Any metric averaged over queries rather than classes is weighted by photographic popularity.
• Identifiers are dataset-local. Statues without a Wikidata item are keyed by a slug of their Commons category title, and a renamed category changes the slug. Pin the revision and the manifest hash below if you need stable keys.
• The files are not EXIF-faithful. The 1,538 re-encoded files lost their EXIF; the 153 untouched ones retain theirs.

Provenance

Built from manifest 3a0bec9daf2c5952. provenance.json records every input hash, the pinned backbone revisions and the digest of every published file.

Citation

Curated by Turhan Can Kargın.

@software{kargin2026krasnalid,
  author    = {Kargın, Turhan Can},
  title     = {Krasnal-ID: fine-grained visual instance retrieval of
               Wrocław's dwarf statues},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22548023},
  url       = {https://doi.org/10.5281/zenodo.22548023}
}

The DOI archives the code and this dataset's build pipeline; the photographs are here. Please also credit the photographers, which credits.csv makes a one-column job.