🐝 Computer Vision for Queen Bee Detection

About this simulation

Real queen-detection systems run an object-detection model (a CNN such as a YOLO variant) over a photo or video frame of the comb, scoring every candidate region and drawing a bounding box wherever the score clears a chosen confidence threshold. This scene renders a simplified 3D hive frame with one queen, a few drones, and a crowd of workers, then simulates that scoring process live so you can see how each pipeline choice changes what the model "sees."

How it works

• Each bee gets a base detection score from features real models learn to weigh — the queen's elongated, tapered abdomen and larger size score highest, drones next, workers lowest.
• Image noise/blur simulates poor lighting, motion blur, or comb occlusion in the source dataset — it randomly perturbs every score, which can both hide the queen and trigger false alarms on look-alike bees.
• Model tier mimics the real deployment trade-off: a lightweight edge model (cheap, runs on-hive in real time) is noisier and less accurate than a larger cloud CNN evaluated on stored frames.
• A box is drawn only when a bee's final score exceeds the confidence threshold — exactly the post-processing step (thresholding, then usually non-max suppression) every detector applies before showing results.

Controls

• Model tier — switch between an edge-deployed and a cloud/full-CNN detector; changes noise sensitivity and simulated latency.
• Confidence threshold — raise it to cut false positives at the risk of missing the queen; lower it to catch more candidates at the risk of more false alarms.
• Image noise/blur — models the effect of dataset and capture quality on detection reliability.
• Bees on frame — more bees means more occlusion and more opportunities for false positives, similar to a denser real frame.
• Show ground-truth box — overlays the true queen location (dashed) so you can compare it against what the model actually reports.
• ↺ New frame — re-randomises bee positions, as if a new photo had been captured. Drag to rotate, scroll to zoom.

Did you know?

Because a colony has only one queen among tens of thousands of workers, queen detection is an extreme class-imbalance problem — labelling teams often have to actively curate "hard negative" drone and cluster crops so the model learns to tell a queen apart from look-alikes rather than just learning "rare thing is probably the queen."