Inspect what the detector found.

Choose a drone photo and change the confidence threshold to see which cattle observations remain. These boxes are saved output from the selected detector, not hand-drawn examples. Moving the slider filters saved predictions; it does not rerun the model.

The live detector runs on the project website in your browser.

Teal boxes: model-proposed cattle. Lower thresholds show more candidates and can also introduce false positives. The count is for this photograph, before cross-photo matching.

What model is this?

YOLOv8n, fine-tuned for one cattle class on ICAERUS drone imagery. The selected checkpoint is corrected-full-yolov8n-v2. Original photos are processed in 1024 × 1024 tiles with 20% overlap, then duplicate tile detections are removed with global NMS at IoU 0.50.

Recorded precision

81.92%

Recorded recall

77.91%

Recorded F1

79.86%

Measured at confidence 0.70 against 1,041 provisional annotations across 77 images. These figures stay fixed when you move the slider: they are recorded benchmark results, not recalculated scores for the selected photo.

Some source annotations omit visible cattle, and the benchmark had already been inspected. Independent field validation is still needed. Detection does not establish complete herd inventory or animal identity.