New Release: Ultralytics v8.3.39

Discussion
,
1

:rocket: Ultralytics Release v8.3.39: Smarter, Faster, Better Models!

We’re thrilled to announce the release of Ultralytics v8.3.39! This update brings significant upgrades to enhance functionality, accuracy, and user experience, making it one of our most impactful releases yet. Below is a summary of the key highlights and improvements.

:star2: Summary

The v8.3.39 release focuses on refining model behavior, improving tooling, and enhancing the overall usability of the platform. With new features, fixed issues, and optimized workflows, this version offers something for everyone—from developers to data scientists. Don’t miss out—explore what’s new below! :star2:

:bar_chart: Key Changes

:brain: Fixed Classification Validation Loss

  • Improved the consistency and accuracy of classification validation.
  • Enhanced the application of softmax operations, limited only to appropriate scenarios, ensuring precise model behavior.

:dart: New “Classes” Argument for Training

  • Added a classes parameter for training configuration, empowering users to train models on selected class IDs only! This allows for focused and highly specific model training workflows.

:movie_camera: Enhanced Video Annotation Tool: “Sweep Annotation”

  • Introducing a dynamic utility that enables users to annotate video data by tracking objects along an interactive sweep line. Great for advanced tracking use cases!

:framed_picture: Improved Color Handling for LibTorch Inference

  • Added BGR to RGB conversion for the C++ LibTorch inference example, significantly improving color accuracy and ensuring seamless compatibility for YOLO results.

:books: Documentation Upgrades

  • Enhanced clarity and accessibility in documentation, including:
    • Clickable YOLO11 performance plots that link directly to docs.
    • High-quality tutorial videos across the platform.
    • Improved readability and consistency in README files.
  • Corrected nomenclature by transitioning all references of YOLOv11 to the proper term, YOLO11.

:gear: Code Refinements

  • Refined segmentation handling by improving out-of-bounds coordinate clipping—fewer errors, smoother results.
  • Simplified attribute access for the Model class using the __getattr__ method—now you can easily access model attributes like stride or task.
  • Improved logging for a more seamless developer experience.

:dart: Purpose & Impact

This update improves accuracy, usability, and functionality across multiple areas:

  • Higher Accuracy: Fixed classification validation scaling issue to deliver better metrics during validation phases.
  • Training Flexibility: The new classes filter lets you pinpoint specific class IDs, optimizing resources and driving focus in model training.
  • Streamlined Annotation: The new sweep annotation tool is a game-changer for video annotation workflows!
  • Better Inference Quality: Proper color handling in LibTorch boosts accuracy for users in C++ environments.
  • Improved User Onboarding: Revamped documentation, clickable visual aids, and tutorial videos make starting with Ultralytics easier than ever!

:clipboard: What’s Changed

Full changelog available here.

:raised_hands: How to Get Started

Try v8.3.39 now and experience improved tools, better performance, and smoother workflows. Whether you’re training, validating, annotating, or deploying, this release has updates just for you!

To upgrade:

pip install ultralytics --upgrade

Check the latest tutorials and documentation at Ultralytics Documentation.

:raised_hand: We’d Love Your Feedback!

Your feedback helps us improve. Let us know your thoughts and experiences with v8.3.39! Share your journey on GitHub Discussions.

Big thanks to our amazing developers and contributors who made this release possible. Explore the changes, and as always, visit our GitHub Repository to contribute or report issues. Let’s continue to innovate, together. :rocket: