phyloforfun/lm3-evaluation-landmark-detector

LM3 Evaluation Dataset: Landmark Detector

250 CC0 single-leaf images cut from herbarium sheets, with 31-keypoint leaf-landmark labels (the mid15_pet5 skeleton: midvein, petiole, apex and base angles, width) for LeafMachine3's landmark detector. Provided as an evaluation dataset for any leaf keypoint model; the LeafMachine3 scores below are one example. Score LM3 with the held-out model: LM3's default landmark model was trained on some of these leaves (see Results).

Task	Keypoint detection (pose), 1 class (leaf), 31 keypoints
Size	250 leaf images, 7,472 labeled keypoints
Splits	complete (201 leaves, all 31 keypoints labeled), incomplete (49 leaves, 10-30 of 31 labeled)
Labels	YOLO pose, labels/<split>/*.txt
Models scored	lm3_landmark_detector__yolo26x_pose_640_canonical__eval_holdout, lm3_landmark_detector__yolo26x_pose_640_canonical
Images license	CC0 1.0, checked per image against GBIF
Labels license	AGPL-3.0, attribution required

Splits: complete and incomplete

Both splits are for evaluation; they divide one set by how completely each leaf is labeled.

• complete: 201 leaves with all 31 keypoints labeled.
• incomplete: 49 leaves where some keypoints are not labeled because the part is absent, hidden, damaged or outside the crop. They are kept on purpose as hard cases. Unlabeled keypoints are written 0 0 0 and are not scored, so these leaves are scored on the keypoints they have. leaves_missing_some_points.json lists the missing keypoints of each leaf.

Report two numbers: all leaves (both splits pooled) and complete leaves only.

Keypoint group	Incomplete leaves missing some or all of it
petiole	25
petiole_tip	25
apex	19
base	15
lamina_tip	11
width	4
midvein	3
lamina_base	3

Split complete

Split incomplete: no petiole is visible, so petiole_0..petiole_4 and petiole_tip are not labeled

Orange = left point, blue = right point (left/right with the leaf tip up), teal = midvein (arrows run from midvein_0 toward the base), violet = petiole, diamond = petiole_tip, dark dots = lamina tip and base and the angle vertices.

Results with LM3 Models

Model	Weights	Trained without these leaves	All: mAP50	All: mAP50-95	Complete: mAP50	Complete: mAP50-95
lm3_landmark_detector__yolo26x_pose_640_canonical__eval_holdout	it3, eval holdout (f90f952)	yes	0.950	0.738	0.955	0.783
lm3_landmark_detector__yolo26x_pose_640_canonical	it3 (736d2c8)	no	0.961	0.790	0.972	0.825

Pose mAP (OKS). lm3_landmark_detector__yolo26x_pose_640_canonical__eval_holdout is LM3's landmark model retrained with the same recipe and these 250 leaves held out (all in its test split); it is in eval_holdout/ of the model repo and is not what LM3 runs. LM3's default (lm3_landmark_detector__yolo26x_pose_640_canonical, the revision pinned by LeafMachine3 3.0.1064) was trained on 63 of these leaves and validated on 11, and its predictions were the starting labels for the other 165 (reviewed and corrected by hand), so its scores here are optimistic.

Scored on CPU with eval/evaluate_onnx.py; eval/evaluate_lm3.py gives identical numbers. Each model's scores for all leaves and for each split are in results/.

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-landmark-detector --repo-type dataset --local-dir lm3-eval-landmark
python lm3-eval-landmark/eval/evaluate_onnx.py                    # all 250 leaves, also each split on its own
python lm3-eval-landmark/eval/evaluate_onnx.py --split complete   # the 201 complete leaves only

The scripts score the held-out model by default; --model all adds LM3's default. 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-landmark/eval/evaluate_lm3.py                    # all 250 leaves
uv run --frozen --no-sync python ../lm3-eval-landmark/eval/evaluate_lm3.py --split complete   # complete leaves only

With Ultralytics:

yolo pose val model=<weights> data=lm3-eval-landmark/data.yaml split=test            # all 250 leaves
yolo pose val model=<weights> data=lm3-eval-landmark/data_complete.yaml split=test   # complete leaves only

Ultralytics letterboxes images in batches, so its numbers differ slightly from the scripts (which letterbox one image at a time, as LM3 does). Scoring is Ultralytics' pose mAP: OKS with sigma 1/31 for every keypoint and area = 0.53 × box area. To score another model, predict 31 keypoints per image in the order below and use kpt_iou and Accumulator from eval/metrics.py.

Contents

images/complete/                 201 leaf images (JPEG) + metadata.csv
images/incomplete/               49 leaf images (JPEG) + metadata.csv
labels/<split>/                  one YOLO pose file per image
data.yaml                        Ultralytics dataset file, all 250 leaves
data_complete.yaml               Ultralytics dataset file, the 201 complete leaves
leaves_missing_some_points.json  the incomplete leaves and the keypoints each one is missing
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, for all leaves and per split
docs/                            the example images on this card
LICENSE.md                       licenses of each part

metadata.csv (per image): GBIF record and link, taxon, institution, catalog number, license, crop size, keypoints labeled and missing, where the label started (label_start), and whether LM3's default model trained or validated on the leaf (lm3_default_model_use).

Label format

Each image is one leaf cut from a herbarium sheet, with a white border of 10% of its width added left and right and 10% of its height added top and bottom: the input LM3's landmark model expects (LM3 adds the same border to every leaf crop). Each label file has one line, normalized 0-1 to the bordered image:

0 cx cy w h  x0 y0 v0  x1 y1 v1  ...  x30 y30 v30

The box is the leaf crop without the border. v is 2 for a labeled keypoint and 0 for a keypoint that is not labeled (x = y = 0). Left and right are relative to the midvein with the leaf tip up. The 15 midvein and 5 petiole points are spaced evenly along the traced line, from the tip toward the base and from the blade toward the free end.

Index	Keypoints
0	lamina_tip
1-3	apex_left, apex_center, apex_right (apex angle: arm, vertex, arm)
4-18	midvein_0 .. midvein_14 (tip to base)
19-21	base_left, base_center, base_right (base angle)
22	lamina_base
23-27	petiole_0 .. petiole_4
28	petiole_tip (free end of the petiole)
29-30	width_left, width_right (leaf width)

flip_idx for horizontal flips: [0, 3, 2, 1, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 21, 20, 19, 22, 23, 24, 25, 26, 27, 28, 30, 29].

How the images were chosen

• Held out from LM3's leaf data. Every leaf is in the test split of LM3_Leaf, LM3's leaf-segmentation data, and these are the same leaves as LM3's leaf-segmentation evaluation set. Other leaves from the same sheets can be in LM3's training data.
• 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 3 LM3 sampling groups and 3 labeling projects, 152 taxa and 18 institutions; one leaf per specimen.

Labels

85 leaves started from LM3's corrected training labels and 165 from predictions of LM3's default landmark model. Every leaf was then reviewed by hand and approved (250 of 250; 187 edited). Finally the midvein and petiole points were respaced evenly along the reviewed line (214 leaves; the traced path is unchanged), the rule LM3's training labels follow.

Limitations

• Small: 250 leaves, so scores move by a few points with a handful of mistakes.
• 165 labels started from LM3's default model. Points that were already right were left in place, which favors that model and models like it.
• Arm points of the apex and base angles and the width ends lie on the leaf margin; where along the margin is partly a labeling choice, so those points cap the reachable OKS.
• Scores cover the landmark model on these crops, not the full LM3 pipeline (where the plant detector makes the crops).

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 images/<split>/metadata.csv):

Images	Institution	URL
85	Smithsonian Institution	https://naturalhistory.si.edu/research/botany
41	Carnegie Museum of Natural History	https://carnegiemnh.org/
40	Naturalis Biodiversity Center	https://www.naturalis.nl/en/collection
20	University of Minnesota	https://www.bellmuseum.umn.edu/
16	Denver Botanic Gardens	http://www.botanicgardens.org/
12	Cheadle Center for Biodiversity and Ecological Restoration	http://ccber.ucsb.edu/
10	University of Tennessee - Knoxville	http://herbarium.utk.edu/
6	Utah State University	http://www.usu.edu
4	National Museum of Natural History, Luxembourg	https://www.mnhn.lu
3	California Academy of Sciences	https://www.calacademy.org/scientists/
3	University of Manitoba	https://umanitoba.ca/
3	Université de Montréal Biodiversity Centre	http://www.irbv.umontreal.ca/recherche/collections/herbier-marie-victorin
2	Appalachian State University	http://herbarium.appstate.edu/
1	Baylor University, Mayborn Museum Complex	https://www.baylor.edu/mayborn/
1	Botanical Research Institute of Texas	http://www.brit.org/herbarium
1	North Carolina Museum of Natural Sciences	https://naturalsciences.org
1	Texas Tech University	http://mossmatters.com/herbarium
1	University of South Carolina	http://www.herbarium.org