Getting Started with YOLOv8: Real-Time Object Detection for Everyone

Introduction

Object detection has been transformed by the YOLO (You Only Look Once) family of models. With YOLOv8, Ultralytics has delivered a framework that is fast, accurate, and incredibly easy to use — whether you are a researcher or an embedded systems engineer.

In this post, I will walk you through training your first YOLOv8 model, from dataset preparation to deployment on edge devices.

Why YOLOv8?

YOLOv8 brings several improvements over its predecessors:

• Anchor-free detection — eliminates the need to hand-craft anchor boxes
• Unified CLI and Python API — yolo train, yolo val, yolo predict
• Modular architecture — Nano (n) to Extra-Large (x) variants for any compute budget
• Multi-task support — detection, segmentation, classification, pose estimation

Installation

pip install ultralytics

That is all. The ultralytics package ships with everything you need.

Training Your First Model

Dataset Structure

YOLOv8 expects the following directory layout:

dataset/
  images/
    train/   # training images
    val/     # validation images
  labels/
    train/   # YOLO format .txt labels
    val/
  data.yaml  # dataset config

data.yaml

path: /path/to/dataset
train: images/train
val: images/val
 
nc: 3  # number of classes
names: ["cat", "dog", "bird"]

Start Training

from ultralytics import YOLO
 
# Load a pre-trained model
model = YOLO("yolov8n.pt")
 
# Train on custom dataset
results = model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    device="cuda",  # or "cpu"
    project="runs/detect",
    name="my_model",
)

Evaluating Performance

After training, evaluate on your validation set:

metrics = model.val()
print(f"mAP50: {metrics.box.map50:.4f}")
print(f"mAP50-95: {metrics.box.map:.4f}")

A mAP50 above 0.85 is generally considered production-ready for most applications.

Inference

results = model.predict(
    source="test_image.jpg",
    conf=0.5,
    iou=0.45,
    save=True,
)
 
for r in results:
    boxes = r.boxes.xyxy  # bounding boxes
    confs = r.boxes.conf  # confidence scores
    cls   = r.boxes.cls   # class indices
    print(boxes, confs, cls)

Exporting for Edge Deployment

One of the most powerful features of YOLOv8 is its export pipeline:

# Export to ONNX for cross-platform inference
model.export(format="onnx", simplify=True)
 
# Export to TensorRT for NVIDIA edge devices
model.export(format="engine", half=True, device=0)
 
# Export to TFLite for mobile / Raspberry Pi
model.export(format="tflite", int8=True)

Performance Benchmarks

Here is a comparison of YOLOv8 variants on a Raspberry Pi 4B (ARM Cortex-A72):

For real-time applications on constrained devices, YOLOv8n with INT8 quantization achieves approximately 18–22 FPS on Raspberry Pi 4.

Conclusion

YOLOv8 is an outstanding starting point for anyone entering computer vision. Its clean API, strong community, and excellent documentation make it the de facto choice for research and production alike.

In upcoming posts, I will cover:

• Structured pruning to reduce model size by 60%
• Deploying to Raspberry Pi with TFLite and OpenCV
• Multi-camera IoT setups with MQTT streaming

Stay tuned!