Ingredient
YOLO (object detection)
Also known as: You Only Look Once, YOLOv8, YOLOv10, Ultralytics YOLO
Family of real-time object-detection neural networks — the dominant choice when a robot needs to find objects (fruit, weeds, animals, equipment) in images and report their bounding boxes and classes at video frame rates. YOLOv8 / v10 / v11 from Ultralytics are the current generation; trained models for general objects (COCO classes) are downloadable, custom training on agricultural datasets is straightforward via Ultralytics's Python API. The right ingredient when computer vision needs to be fast, accurate, and deployable to edge devices like a Jetson. Apache 2.0 (early) → AGPL (current) — license matters for commercial deployments.
Inputs / outputs
- Input: RGB image, typical resolutions 320×320 to 1280×1280
- Output: array of bounding boxes — each with class ID, confidence, x/y/w/h
- Speed: YOLOv8-nano: 8–15 ms per frame on [[nvidia-jetson|Jetson Orin Nano]]; 50–100 ms on Pi 5
- Models: nano, small, medium, large, X-large — pick by speed/accuracy tradeoff
Solves / unlocks
- Real-time fruit detection for picking ([[strawberry|strawberry]], tomato, apple)
- Weed-vs-crop differentiation in rows for targeted spraying
- Pest detection on sticky traps (count whiteflies, aphids, thrips)
- Livestock counting and identification
- Equipment-and-worker safety detection (people in robot path)
- Crop-row tracking (detect row endpoints, plan turn-arounds)
Constraints
- AGPL license since YOLOv5+ — commercial use requires Ultralytics commercial license; budget if shipping a product.
- Training data is the work — YOLO is the easy part; collecting and labeling 1000+ images per class is the bottleneck.
- Generalization gap — model trained on one farm often fails on another (lighting, crop variety, growth stage); plan for retraining.
- Small objects are hard — tiny pests, small seedlings need higher input resolution or explicit small-object training.
Source
- Ultralytics: https://github.com/ultralytics/ultralytics (AGPL or commercial)
- YOLOv7 (older, GPL-3): https://github.com/WongKinYiu/yolov7
- YOLO-NAS (Apache 2.0 alternative): https://github.com/Deci-AI/super-gradients
- Roboflow (dataset + training UI): https://roboflow.com/
See also
Auto-generated from this entry’s typed relations: frontmatter, grouped by relation type so the editorial signal isn’t flattened.
- Member of: [[ingredient]]
- Combines with: [[opencv]] · [[nvidia-jetson]] · [[raspberry-pi]] · [[rgbd-camera]] · [[ros2]] · [[plantvillage-dataset]]
What links here, and how
Inbound connections from across the wiki, grouped by lens and by relationship. These appear automatically — every entity page declares what it links to, and that data populates here on the targets.
Practical
contains
- Farm-tech toolkit framework / real-time object-detection neural network
combines with
- iNaturalist API iNaturalist as labeled-image source for training custom YOLO weed/beneficial detectors
- PlantVillage dataset YOLO trained on PlantVillage for combined leaf-detection + disease-classification in single model
combines
- Recipe: autonomous row-crop weeder custom-trained model: weed vs crop bounding boxes per frame
4 inbound links · 7 outbound