Ultralytics YOLOv5 v6.0 is here!

Our latest update arrived on October 12th, 2021 and is the first major release since April 2021. Release v6.0 brings significant improvements that lower memory requirements during training, increase accuracy during deployment, and optimize runtime performance across the full range of Ultralytics YOLOv5 models.

Ultralytics YOLOv5 v6.0 speed vs. accuracy plot.

The result for ML engineers and data scientists is that YOLOv5 now provides a more powerful Vision AI solution, and is even easier to train and deploy than ever before. Multiple updates to the model backbones were made based on empirical results of Ultralytics R&D efforts.

Modifications include new modules and improvements to existing modules that combine to produce faster, smaller, and more accurate models.

We couldn’t have done this by ourselves though! This release incorporates 465 PRs from 73 contributors from around the world, all collaborating together to push the boundaries of AI. See our open-source contributing guidelines if you’d like to learn more or contribute yourself.

This release brings hundreds of small changes that accumulate to make a real difference, far too many to go into detail, but a few of the major highlights are:

• Roboflow Integration ⭐ NEW: Train YOLOv5 models directly on any Roboflow dataset with our new integration! This integration provides a direct connection between your Roboflow datasets and your YOLOv5 trainings. (#4975 by @Jacobsolawetz)
• YOLOv5n 'Nano' models ⭐ NEW: New smaller YOLOv5n (1.9M params) model below YOLOv5s (7.5M params), exports to 2.1 MB INT8 size, ideal for ultralight mobile solutions. (#5027 by @glenn-jocher)
• TensorFlow and Keras: TensorFlow, Keras, TFLite, TF.js model export now fully integrated into YOLOv5 for smoother transitions from training to deployment. (#1127 by @zldrobit)
• OpenCV DNN: YOLOv5 ONNX models are now compatible with both OpenCV DNN and ONNX Runtime to provide users with even more deployment destination options. (#4833 by @jebastin-nadar)
• Model Architecture: Updated backbones are slightly smaller, faster and more accurate, and require less GPU memory during training.

Final Thoughts#

A little over a year after releasing YOLOv5, our state-of-the-art object detection technology is now on its way to becoming the world’s most loved vision AI. With the help of hundreds of collaborators and feedback from thousands of users, we are creating tools that are both effective and easy to use, and our new v6.0 release is the next exciting step in this journey. Head on over to our open-source GitHub repository to get started using YOLOv5 today! YOLOv5 GitHub Repository