Tutorial 2 — Train a Custom Defect Detection Model with YOLOv8

The pre-trained COCO model detects everyday objects. In this tutorial you will train your own model to detect the specific defects or objects that matter for your use case — scratches, cracks, missing components, or anything else you need to find.

What you will learn

• How to collect and annotate your own image dataset
• How to structure a YOLOv8 training dataset (YAML format)
• How to train on Google Colab (free GPU) and export to TensorRT
• How to deploy your custom model on Jetson and run inference
• How to evaluate model accuracy with mAP metrics

Step 1 — Collect your images

Aim for at least 200 images per defect class. Use the included capture script to take photos with the Jetson camera and save them automatically:

cd ~/tutorials/02-custom-defect-detection
python3 capture_images.py --output ./dataset/images/train --count 200

Step 2 — Annotate with Label Studio (pre-installed)

# Start Label Studio annotation tool (pre-installed on your kit)
label-studio start --port 8080

Open http://localhost:8080 in your browser. Import your images, draw bounding boxes around each defect, label them, and export in YOLO format.

Step 3 — Create your dataset config

# dataset/data.yaml
path:  /home/user/tutorials/02-custom-defect-detection/dataset
train: images/train
val:   images/val

nc: 2   # number of classes
names: ['scratch', 'crack']   # your class names

Step 4 — Train on Google Colab (free GPU)

# Run this in Google Colab (free T4 GPU)
!pip install ultralytics

from ultralytics import YOLO
model = YOLO("yolov8s.pt")    # start from pretrained weights

model.train(
    data    = "data.yaml",
    epochs  = 100,
    imgsz   = 640,
    batch   = 16,
    device  = 0
)
# Best weights saved to: runs/detect/train/weights/best.pt

Step 5 — Export to TensorRT and copy to Jetson

# In Colab — export to TensorRT engine
model.export(format="engine", half=True, imgsz=640)
# Downloads: best.engine

# On Jetson — copy and run
python3 detect.py --model ~/my_models/best.engine --source 0 --show

✅ Next: Tutorial 3 — Pose Estimation | Back to Jetson Kit