YOLOv8 Fine-Tuning
By Author • September 15, 2025
We often explore how computer vision can solve real-world problems. One exciting experiment we carried out was fine-tuning YOLOv8, the latest version of the popular “You Only Look Once” (YOLO) model, to recognize car brands.
Here’s how we approached the challenge, what we learned, and why it matters for businesses that want smarter AI-powered recognition systems.
Why YOLOv8?
Traditionally, fine-tuning deep learning models has been complex, requiring large datasets, tricky configurations, and powerful hardware. YOLOv8 changes that.
With a simple folder structure and built-in support for transfer learning, it allows even smaller datasets to achieve strong results. This makes it practical, resource-efficient, and beginner-friendly — perfect for businesses that want results without massive infrastructure.
Preparing the Custom Dataset
We trained YOLOv8 to recognize five major brands: BMW, Audi, Mercedes-Benz, Toyota, and Jaguar. Each brand had images taken from different angles and lighting conditions, so the AI could learn brand-specific features like logos, grilles, and car shapes.
To make the system more realistic for real-world use, we added an “Others” class — so cars outside these five brands wouldn’t be wrongly classified. We also included a “No Car” detection, ensuring the model wouldn’t try to classify irrelevant images (like a building or tree).
Transfer Learning: Building on What’s Already Known
Instead of training from scratch, we used a pre-trained YOLOv8 model (yolov8x-cls.pt). This model already understood basic visual features like shapes, edges, and textures. We simply fine-tuned it to specialize in distinguishing between car brands.
The benefits were clear:
- Faster training time
- Higher accuracy with less data
- More efficient use of resources
Training the Model: From Early Struggles to Reliable Results
- Early stages (0–40 epochs): The AI was learning but often misclassified cars, especially at unusual angles or with busy backgrounds.
- Mid stages (40–80 epochs): Accuracy improved steadily. The model began recognizing brand features like BMW grilles or Audi logos.
- Final results (80 epochs): The system consistently classified the five brands with strong accuracy, even in challenging conditions.
Making the Model Smarter: Hyperparameters and Monitoring
Behind the scenes, fine-tuning involved adjusting “hyperparameters” (training settings) and carefully monitoring progress:
- Batch size: 16 worked best — larger caused memory issues, smaller slowed training.
- Augmentation: Exposing the model to flips, rotations, and lighting changes improved flexibility.
- Image size: 224×224 was the sweet spot — detailed enough for accuracy without overloading the system.
- Checkpoints: Saved progress during training, so interruptions didn’t mean starting over.
Real-World Impact
When we tested the fine-tuned model:
- It accurately recognized the five main car brands across different angles and environments.
- It classified unknown brands as “Others” instead of forcing a wrong answer.
- It rejected non-car images, responding with “No car found.”
- These improvements turned the project from a neat demo into a practical solution ready for deployment.
Key Takeaways
Our YOLOv8 fine-tuning project showed us:
- Simplicity can be powerful. A clean dataset structure and transfer learning make advanced AI accessible.
- Accuracy isn’t enough. Robust handling of “unknown” and “no car” cases is what makes AI usable in the real world.
- Monitoring is essential. Real-time insights and checkpoints keep long training runs efficient and reliable.