Evaluation complete · launch ready

Model report · YOLO26s

EleFind YOLO26s

A 9.95-million-parameter, single-class elephant detector fully fine-tuned for nadir aerial photography. Its launch evaluation uses held-out source photographs, metre-based point matching, altitude-aware box analysis, and a direct YOLO11s baseline comparison.

0.9002

Point AP

2.25 m tolerance

0.8896

Best F1

confidence 0.50

0.9168

Precision

at best F1

0.8640

Recall

at best F1

Point AP is the primary metric because the source dataset provides elephant centre points rather than measured ground-truth boxes. All headline test results use cross-tile NMS/IOU after tiled inference on held-out full-resolution photographs.

Evaluation context

How this model was measured

439

Full-resolution test images

438 annotated source photos plus one negative image, held out from model training.

2,970

Test elephant points

Centre-point annotations scored at 1.0 m, 2.25 m, and 4.5 m matching tolerances.

17,385

Training tiles

5,795 positive and 11,590 background tiles derived from 1,304 source photos.

9.95 M

Parameters

A compact 19.5 MB model with all 902 layers transferred for full fine-tuning.

Operating configuration

Required slice size: 1024 × 1024 px
Required overlap: 0.30
Inference confidence: 0.30
Best-F1 confidence: 0.50
Cross-tile post-process: NMS / IOU
Post-process threshold: 0.40
RTX 3060 · 20 MP: ≈3.2 s
CPU · 20 MP: ≈14.4 s

Evaluation figures

What the model learned

Point-based precision–recall
Primary evaluation at three metre-based centre-matching tolerances; AP reaches 0.9002 at 2.25 m.

Confidence sweep
Precision, recall, and F1 across confidence thresholds. Best F1 is 0.8896 at confidence 0.50.

Altitude-aware predicted boxes
Predicted box width follows the elephant size implied by each photograph's ground sampling distance.

Synthetic-box metric sensitivity
IoU-based scores change materially when different invented box sizes are wrapped around point annotations.

At the 2.25 m operating tolerance, the detector records 2,566 true positives, 233 false positives, and 404 false negatives.

Direct full-image inference is not supported: recall drops to roughly 0.1% when a full survey image is resized instead of tiled.

NMS/IOU is the recommended cross-tile merge rule. The default greedy merge produced lower point AP and lower maximum recall in this evaluation.

Model comparison

YOLO11s baseline vs YOLO26s

Both models below were evaluated on the same held-out source photographs using the same tiled pipeline and metre-based matching. YOLO26s makes a modest point-metric gain, while its adaptive predicted boxes produce a much stronger result under the 4.5 m synthetic-box convention.

Point AP @ 2.25 msame held-out evaluation

YOLO11s baseline

0.8962

YOLO26s

0.9002

Best F1same held-out evaluation

YOLO11s baseline

0.8854

YOLO26s

0.8896

Synthetic-box mAP@0.5same held-out evaluation

YOLO11s baseline

0.1221

YOLO26s

0.8584

Evaluation-rule comparison
Point-distance scoring remains stable across tolerances, while box mAP changes sharply when different synthetic box conventions are applied to the same source annotations.

Known limits

Tiled inference is required for full-resolution survey photographs.
The evaluated altitude range corresponds to 0.024–0.130 m per pixel.
Domain shift beyond aerial savanna and bushland can reduce performance.
Outputs require human verification before census or enforcement use.

Try detection Read the YOLO11 report