New Release: Ultralytics v8.3.76

Discussion
,
1

:rocket: Announcing Ultralytics v8.3.76: Enhanced Dynamic Inference, Tracking, and Documentation :star2:

We’re excited to share the release of Ultralytics v8.3.76, packed with significant improvements to model deployment, tracking, and user-friendliness. This update directly addresses community feedback and introduces enhancements to boost efficiency and developer ease. Let’s dive into the highlights! :tada:

:star2: Summary

This release introduces:

  • Improved dynamic batch inference for ONNX models (with NMS export).
  • A smoother object tracking experience with key fixes for tensor-based workflows.
  • Better documentation and examples to help you navigate and utilize these features seamlessly.

:bar_chart: Key Changes

Dynamic Batch Improvements

  • Resolved issues with dynamic=True and nms=True in ONNX exports, allowing dynamic batch size support.
  • Introduced input padding during ONNX export to handle varying batch sizes robustly.

Tracking Enhancements

  • Fixed errors when working with torch tensors in model.track().
  • Integrated better handling of input images with trackers, improving accuracy for tracked objects.

Performance Accuracy

  • Corrected GPU memory conversion errors for accurate VRAM usage logging.

Documentation Updates

  • Standardized formatting across the documentation for clarity.
  • Added examples for interpreting predictions across tasks like detection, pose estimation, segmentation, and more.

Other Code Refinements

  • Adjusted log details to display layers without parameters for transparency in the architecture reporting.
  • Improved GitHub issue templates for smoother reporting and handling of feature requests vs. bug reports.

:dart: Purpose & Impact

  • :hammer_and_wrench: Enhanced Deployment: Dynamic input padding ensures seamless ONNX model experiments and deployments.
  • :movie_camera: Tracking Upgrades: Developers and end-users benefit from consistent and reliable tracking output.
  • :clipboard: Accurate Debugging: Clearer VRAM usage metrics bolster debugging and optimize memory usage.
  • :books: Developer Empowerment: Refined docs and practical examples make features more user-friendly.
  • :rocket: Streamlined Efficiency: Optimized architecture leads to faster and more reliable handling of models.

:new: Detailed Changes

Here’s the list of PRs included in this release:

:scroll: View the full changelog here: compare/v8.3.75…v8.3.76.

:tada: Special Thanks

A warm welcome to @LoveAndHope-dev, who made their first contribution in this release! Thank you for being a valuable part of the YOLO community.

:bulb: Try It Out!

We encourage all users to explore the new features in v8.3.76. Whether you’re deploying models in ONNX, refining tracking pipelines, or diving into the updated documentation, we’re excited to hear your thoughts.

Feel free to open a discussion or report issues on our GitHub repo. Your feedback is invaluable in shaping future releases. :pray:

:arrow_right: Download the latest release here: v8.3.76.

Thank you for being part of the YOLO journey, and happy experimenting! :blush: