phyloforfun/lm3-evaluation-archival-detector

LM3 Evaluation Dataset: Archival Detector

100 CC0 herbarium sheet images with box labels for the 9 non-plant objects LeafMachine3's archival detector finds: rulers, barcodes, color cards, labels, maps, envelopes, photos, attached items and weights. Provided as an evaluation dataset for any object detector; the LeafMachine3 scores below are one example. No LM3 model was trained on these images.

Task	Object detection, 9 classes
Size	100 images, 1,116 labeled objects
Labels	YOLO boxes, labels/test/*.txt
Models scored	lm3_archival_detector__yolo26x_det_1280, lm3_archival_detector__yolo26n_det_640
Images license	CC0 1.0, checked per image against GBIF
Labels license	AGPL-3.0, attribution required

Results with LM3 Models

LM3's own models are evaluated on a larger test set (LM3_Archival, 875 sheets) that also contains non-CC0 images, which cannot be published on Hugging Face. These are the scores on this CC0 subset.

Model	Role in LM3	Weights	mAP50	mAP50-95
lm3_archival_detector__yolo26x_det_1280	default	it1 (519b73c)	0.982	0.818
lm3_archival_detector__yolo26n_det_640	alternate	it2 (cb621ab)	0.973	0.802

Model revisions are the ones pinned by LeafMachine3 3.0.1038. Scored on CPU with eval/evaluate_onnx.py; eval/evaluate_lm3.py gives identical numbers. Per-class scores are in results/.

Ground truth (left) and prediction (right) for lm3_archival_detector__yolo26x_det_1280 at LM3's default confidence (0.25); all pages for every model are in results/contact_sheets/<model>/images/.

Evaluate models with these labels

Pure Python (ONNX Runtime, no torch):

pip install numpy opencv-python-headless onnxruntime huggingface_hub
hf download phyloforfun/lm3-evaluation-archival-detector --repo-type dataset --local-dir lm3-eval-archival
python lm3-eval-archival/eval/evaluate_onnx.py              # LM3 default model; --model all for every model

With LeafMachine3 (its own inference code):

git clone https://github.com/Gene-Weaver/LeafMachine3 && cd LeafMachine3
uv sync --frozen --extra cpu                        # or --extra gpu / --extra macos
uv run --frozen --no-sync python ../lm3-eval-archival/eval/evaluate_lm3.py

data.yaml also works with Ultralytics: yolo val data=data.yaml split=test.

Contents

images/test/  100 herbarium sheet images (JPEG)
labels/test/  one YOLO file per image: `class cx cy w h`, normalized 0-1
data.yaml     class names (Ultralytics format)
metadata.csv  per image: GBIF record and link, taxon, institution, catalog number, license
SOURCES.csv   the GBIF datasets the images come from, with DOIs
eval/         evaluation scripts and `models.json` (the models and revisions scored)
results/      scores per model (with per-class), and `contact_sheets/<model>/images/` of predictions
LICENSE.md    licenses of each part

How the images were chosen

• Held out from LM3. Every image is from the test split of LM3_Archival, which LM3 models never train on. This set is the CC0 part of that split, the part that can be redistributed.
• CC0 only. The GBIF occurrence record and every licensed image on it were CC0 at the GBIF API (2026-10-12).
• Spread out. Balanced across 7 LM3 sampling groups and 34 labeling projects, 89 taxa and 23 institutions; one image per specimen; sheets with rare classes picked first.
• Reviewed. 1 image was replaced by hand (not a suitable image).

Labels

Drawn for LM3 training, then every image reviewed again for this release: 100 of 100 approved, 1 corrected.

Class	Objects	Images
Ruler	135	99
Barcode	79	77
Colorcard	138	100
Label	706	100
Map	3	3
Envelope	41	39
Photo	1	1
Attached Item	13	11
Weights	0	0

Limitations

• Small: 100 images, so scores move by a few points with a handful of mistakes.
• Rare classes have very few examples (Map 3, Photo 1, Attached Item 13), so their AP is noisy; Weights does not occur in this set.
• Scores cover this model on its own, not the full LM3 pipeline.

License and citation

Images: CC0 1.0. Labels and evaluation code: AGPL-3.0 © 2026 William N. Weaver; attribution required. See LICENSE.md.

Weaver, W. N. (2026) LeafMachine3. https://leafmachine.org

Images courtesy of these institutions (via GBIF; per-image records in metadata.csv):

Images	Institution	URL
20	Carnegie Museum of Natural History	https://carnegiemnh.org/
20	Smithsonian Institution	https://naturalhistory.si.edu/research/botany
15	University of Minnesota	https://www.bellmuseum.umn.edu/
6	Denver Botanic Gardens	http://www.botanicgardens.org/
4	Appalachian State University	http://herbarium.appstate.edu/
4	University of Tennessee - Knoxville	http://herbarium.utk.edu/
3	Cheadle Center for Biodiversity and Ecological Restoration	http://ccber.ucsb.edu/
3	Institute of Plant and Animal Ecology (IPAE)	https://ipae.uran.ru/
3	National Museum of Natural History, Luxembourg	https://www.mnhn.lu
3	Texas Tech University	http://mossmatters.com/herbarium
3	Université de Montréal Biodiversity Centre	http://www.irbv.umontreal.ca/recherche/collections/herbier-marie-victorin
3	Utah State University	http://www.usu.edu
2	California Academy of Sciences	https://www.calacademy.org/scientists/
2	North Carolina Museum of Natural Sciences	https://naturalsciences.org
2	University of Lynchburg	https://www.lynchburg.edu/academics/majors-and-minors/biology/ramsey-freer-herbarium/
1	Audubon Society -- Silver Bluff Audubon Center and Sanctuary
1	Campbell University
1	Muhlenberg College
1	University of Central Arkansas
1	University of South Carolina Salkehatchie	http://herbarium.biol.sc.edu/floracaroliniana/collections/index.php
1	Weymouth Woods Sandhills Nature Preserve	http://www.ncparks.gov/weymouth-woods-sandhills-nature-preserve
1	Yale Peabody Museum	http://www.peabody.yale.edu/