Apiarist Honey Bee Detector

--- license: apache-2.0 library_name: ultralytics tags: - yolo - yolov8 - object-detection - bees - beekeeping - computer-vision pipeline_tag: object-detection --- # Apiarist Honey-Bee Detector (YOLOv8s) A custom-trained **YOLOv8s** specialist detector for honeycomb frame inspection. Built as part of [Apiarist](https://huggingface.co/spaces/build-small-hackathon/Apiarist), a fully-offline AI hive inspector for backyard beekeepers, made for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon). ## Classes Trained on the [hendricks_ricky bee-project](https://universe.roboflow.com/hendricks_ricky-hotmail-de/bee-project) dataset (3,308 labeled images, including 892 queen-bee examples). Four classes: - `Queen`, queen bee (larger, elongated abdomen) - `Worker`, worker / forager bees (the majority class) - `Drone`, male / drone bees - `Varroa`, varroa destructor mites visible on bees or comb ## Usage ```python from ultralytics import YOLO from PIL import Image model = YOLO("honey_bee_detector.pt") results = model(Image.open("hive_frame.jpg"), conf=0.10, device="cpu") for box in results[0].boxes: cls = model.names[int(box.cls.item())] conf = float(box.conf.item()) print(f"{cls}: {conf:.0%}") ``` Recommended per-class confidence thresholds (tuned empirically on real frame photos): ```python PER_CLASS_CONF = { "Worker": 0.25, "Drone": 0.55, # higher, drones false-positive on fingers, shadows "Varroa": 0.30, # mites are small "Queen": 0.15, # lower, try harder to find her } ``` ## Training - Base weights: `yolov8s.pt` - Epochs: 60 - Image size: 640×640 - Batch size: 32 - Hardware: 1× NVIDIA T4 (Modal) - Cost: ~$0.40 of free hackathon credit ## Limitations The training set includes ~892 queen images, which is far more than the typical bee-detection dataset but still limited compared to the millions of bees in worker class. Queen detection precision is moderate, when the model labels a queen, verify visually. Use as a **candidate flag**, not as ground truth. Varroa mite detection works on close-up frames where mites are visible on the dorsal side of bees or in cells, but will miss mites hidden under abdomens. ## License Apache 2.0. Trained on data released under CC BY 4.0 by the original dataset authors. ## Citation If you use this model, please credit: - This work: Apiarist (Build Small Hackathon entry, 2026) - Training d...