Back to EleFind
Deployed research model

Model report · YOLO11

EleFind YOLO11

The first EleFind aerial elephant detector combines a single-class YOLO11 model with tiled inference for high-resolution drone imagery. This report preserves the complete published 50-image evaluation and places it beside the newer, larger comparison study.

84.3%
mAP@0.5
50-image test set
53.2%
Precision
published test result
49.1%
Recall
published test result
51.0%
F1 score
precision–recall balance

These headline values come from the original model's 50-image test set. They are self-reported research results, not a guarantee of field performance. The model-to-model comparison below uses a separate held-out evaluation with a retrained YOLO11s fixed-box baseline.

Evaluation context

How this model was measured

50
Test images

The original evaluation set used for the published precision, recall, F1, and mAP values.

185 / 163 / 192
TP / FP / FN

Detected elephants, false detections, and missed elephants at the reported operating point.

1 class
Elephant detector

Fine-tuned for elephants in overhead aerial and drone photographs.

100
Training epochs

Early stopping used a patience of 20 epochs, with automatic mixed precision enabled.

Operating configuration

Input and slice size
1024 × 1024 px
Slice overlap
0.30
Confidence threshold
0.30
Cross-tile NMS IoU
0.40
Training batch size
16
Initial learning rate
0.01
Augmentation
Mosaic · erase · flip
Intended view
Overhead / nadir
Evaluation figures

What the model learned

Training progression
Training progression
Box and classification loss alongside precision, recall, and mAP over the full training run.
Normalised confusion matrix
Normalised confusion matrix
Shows correct elephant detections as well as confusion with the background class.
Precision–recall curve
Precision–recall curve
The area under this curve produces the reported 0.843 mAP@0.5 result.
The high mAP@0.5 shows that the detector can rank useful candidates, while the lower operating-point precision and recall show that thresholding and review still matter.
Tiling is part of the model workflow: shrinking an entire high-resolution survey frame would make already-small elephants even harder to resolve.
The test set is small. Treat these numbers as an initial research result and use the larger held-out comparison below for model-generation comparisons.
Model comparison

YOLO11s baseline vs YOLO26s

For an apples-to-apples generation comparison, YOLO26s was evaluated beside a separately trained YOLO11s fixed-box baseline on the same held-out Aerial Elephant Dataset photos. That comparison is distinct from the 50-image YOLO11 report above.

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
Point scoring is more stable than invented boxes
Point scoring is more stable than invented boxes
The same detections produce very different IoU scores when synthetic box size changes. Metre-based centre matching stays aligned with the source point annotations.

Known limits

  • The original headline metrics use only 50 test images.
  • Dense vegetation and heavy occlusion can hide small elephants.
  • Rocks, shadows, and similarly sized objects can create false positives.
  • Side-angle imagery is outside the overhead-view training objective.
Try detection Read the YOLO26 report