Object Detection With YOLO: From Dataset to Deployment

The short answer

YOLO (You Only Look Once) is a family of models that find objects in an image in a single pass and draw boxes around them with a class and confidence score. For a business task — a helmet on a worker, a license plate, a product on a shelf — you usually do not train from scratch. You fine-tune a pretrained model on your own data. The path is: data → labeling → training → evaluation → deployment.

Step 1. Collecting images

Model quality depends on how closely your data matches real operation.

• Shoot with the same camera and angle the model will use in production.
• Cover different lighting: day, night, glare, rain.
• Include hard frames: partly hidden objects, crowds, small objects far away.
• Add frames with no objects so the model produces fewer false alarms.
• Avoid near-identical consecutive video frames: take every Nth frame.

Step 2. Labeling

Each object gets a box and a class. Tools such as CVAT or Label Studio export labels in YOLO format.

Rules that save weeks:

• Write a labeling guide: what counts as an object, how to handle partly visible ones.
• Boxes should fit the object tightly, without extra background.
• Spot-check labels — labeling errors directly reduce quality.
• Split data into train / val / test up front, and keep frames from one scene in the same split.

A YOLO dataset description:

path: datasets/helmets
train: images/train
val: images/val
names:
  0: helmet
  1: no_helmet

Step 3. Training

With the Ultralytics library, basic fine-tuning takes a few lines:

from ultralytics import YOLO

model = YOLO("yolov8n.pt")          # small pretrained model
model.train(data="data.yaml", epochs=100, imgsz=640)
metrics = model.val()                # evaluate on val
model.export(format="onnx")          # export for deployment

Model size is a trade-off: small variants (n, s) are fast and fit edge devices, large ones (l, x) are more accurate but need a strong GPU.

Step 4. Evaluation: what mAP means

• Precision — the share of detections that are actually correct.
• Recall — the share of real objects the model found.
• mAP@0.5 — average precision when boxes overlap by 50% or more.
• mAP@0.5:0.95 — a stricter metric averaged across several overlap thresholds.

Look beyond the overall mAP: check per-class metrics and review errors visually. For safety tasks, recall often matters more (do not miss a violation); for automated penalties, precision matters more (do not punish by mistake).

Step 5. Deployment

To speed things up, export the model to ONNX, TensorRT or OpenVINO depending on the hardware. Then you need code around the model: reading the RTSP stream, confirming detections across several frames, and sending events to a CRM, Telegram or an accounting system.

Common mistakes

• Training on internet photos but deploying on a camera with a different angle.
• Evaluating on frames very similar to training ones — the metrics come out inflated.
• Not monitoring after launch: lighting or the camera changes and quality drops.
• Reacting to every single detection without confirming it over several frames.