Skip to main content

Older YOLO Isn't Always Worse: A YOLOv5s vs YOLOv8s Comparison

· 2 min read
Yi-Chen Zhang
Yi-Chen Zhang
Lead Engineer, AI and Autonomous

When people think about YOLO model versions, the assumption is usually straightforward: newer is better. But that isn't always the case — especially when you're using pre-trained weights without any fine-tuning.

The Experiment​

I recently ran a side-by-side comparison between YOLOv5s (top) and YOLOv8s (bottom), both using their official pre-trained weights on the same video footage — no retraining, no fine-tuning.

What I Observed​

By visual inspection alone, YOLOv5s outperforms YOLOv8s on this footage in three noticeable ways:

  • More detections — YOLOv5s identifies more objects in the scene
  • Fewer ghost objects — YOLOv8s produces more false positives
  • Better bounding box sizing — YOLOv5s predicts more accurate bounding box dimensions

Why Does This Happen?​

This might seem counterintuitive, but there are a few reasons why an older model version can outperform a newer one out-of-the-box:

  1. Different training data distributions — Each YOLO version is trained on slightly different data pipelines and augmentation strategies. A newer model optimized for certain benchmarks may not generalize as well to your specific domain without fine-tuning.

  2. Architecture trade-offs — Newer architectures often prioritize efficiency or generalization improvements that show up in aggregate metrics (like mAP on COCO), but may regress on specific scene types.

  3. Pre-trained weight quality — The quality and diversity of the pre-trained checkpoint matters as much as the architecture itself.

The Takeaway​

Fine-tuning and retraining on domain-specific data is always the right path for optimal performance. But if you're in a pinch and need a model that works acceptably out-of-the-box, don't dismiss older YOLO versions just because a newer one exists. It's worth running a quick comparison on your own data before committing to a specific version.

Sometimes the tried-and-true solution is good enough — and occasionally, it's actually better.