Train YOLO on detections obtained from YOLO or from YOLO-World about ultralytics

[github-actions]

👋 Hello @DouwedeKok, 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 and for your kind words about our models!

Regarding your query on training a YOLO model using detections from an untrained YOLO model, this approach can indeed be useful, especially in iterative training or active learning scenarios. However, as you've pointed out, using detections from the same model that generated them without introducing new annotations might reinforce the existing biases or missed detections.

Using YOLO-World, which is designed for open-vocabulary detection, could potentially provide a broader range of detected objects and might help in reducing biases when used as a pre-training step. This could be beneficial before fine-tuning your model on specific tasks like cow detection.

As for the number of detections per class, having 2000+ detections can indeed help in achieving robust performance, but this number can vary based on the complexity of the objects and scenes involved.

Training on detections from a model like YOLO-World and then fine-tuning on your specific dataset might offer a more effective approach than reusing detections from the same model iteration.

I hope this helps clarify your questions!

[DouwedeKok]

Thank you for the response!

Then my assumption is indeed correct. In that case, I will use this approach. Much appreciated 🚀!

[glenn-jocher]

You're welcome! I'm glad to hear that the information was helpful. If you have any more questions as you proceed, feel free to reach out. Happy modeling! 🚀