yoloV8 has lower inference accuracy than yoloV5 about ultralytics

[github-actions]

👋 Hello @tonycc521, thank you for your interest in Ultralytics YOLOv8 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the vibrant Ultralytics Discord 🎧 community for real-time conversations and collaborations. This platform offers a perfect space to inquire, showcase your work, and connect with fellow Ultralytics users.

Install

Pip install the ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8.

pip install ultralytics

Environments

YOLOv8 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLOv8 Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

Hello!

Thank you for reaching out with your observations. It's interesting to note the difference in accuracy between YOLOv8 and YOLOv5 under similar conditions.

There could be several reasons for this behavior:

1. Model Configurations: YOLOv8 includes various enhancements and might require tuning specific parameters differently compared to YOLOv5 for optimal performance on some datasets.
2. Hyperparameters: Default hyperparameters for both versions might not be equally suited for all types of data. Consider customizing the learning rate, batch size, or other training-specific parameters.
3. Feature Enhancements: YOLOv8 is designed with additional features that might be affecting its inference behavior differently.

As a starting point, you might want to examine and adjust the hyperparameters specific to YOLOv8 or even try using a technique like transfer learning from a YOLOv5 model that worked well.

If the issue persists, would you be able to provide more details on the metrics used for evaluating accuracy and any specific configurations used during the training? This information could be helpful in pinpointing the cause of the discrepancy.

Looking forward to helping you get the most out of YOLOv8! 🚀

[tonycc521]

Hello! How does the choice of model size relate to the datasets? How to adjust the learning rate and batchSize?

[glenn-jocher]

Hello!

Great questions! The choice of model size should generally align with the complexity and volume of your dataset. Larger datasets or those with more complex features might benefit from bigger models like YOLOv8x, while smaller, simpler datasets might be efficiently handled by YOLOv8n.

For adjusting the learning rate and batchSize, a general rule is:

• Learning rate: Start with a default (e.g., 0.01) and decrease it if the training process oscillates too much.
• Batch size: Increasing might speed up training but requires more memory. Reduce it if you encounter memory issues.

You can easily adjust these in the training command as follows:

yolo train data=your_data.yaml model=yolov8n.yaml epochs=100 imgsz=640 lr0=0.01 batch=16

Adjust these values based on your specific data characteristics and hardware capabilities.

Hope this helps! Let's get your model training optimized! 🚀

[tonycc521]

When training the model, the map50-90 data has been very low, only about 0.2, 0.3, how should I adjust it?

[glenn-jocher]

Hello!

It sounds like your model might need some hyperparameter adjustments to improve the mAP scores. Here are a couple of tips:

1. Check the Dataset: Ensure your dataset is correctly labeled and diverse enough to cover different scenarios the model might encounter.
2. Modify Learning Rate: If it's too high, the model might not be converging well. Try reducing it slightly.
3. Increase Training Epochs: Sometimes, more training time is required for the model to learn effectively.

You could start experimenting with these parameters like so:

yolo train data=your_dataset.yaml model=yolov8n.yaml epochs=150 lr0=0.005

Keep an eye on the training loss and validation metrics to guide further tweaks. Good luck! 🚀

[tonycc521]

Hello! How are the big targets defined in the datasets? How are the small targets defined in the datasets? Can I change the image size to 1280 for high resolution pictures?

[glenn-jocher]

Hello!

In YOLO datasets, target sizes are often relative to the image size:

• Small targets might be those occupying a small portion of the image (few pixels).
• Big targets are those occupying a significant portion of the image.

Yes, you can definitely increase the image size to 1280 for high-resolution pictures to help the model better recognize details, especially useful for small targets. Here’s how you can set it in your training command:

yolo train data=your_dataset.yaml model=yolov8n.yaml imgsz=1280

Adjusting image size like this can improve detection accuracy for different target sizes. Happy modeling! 🚀