New Release: Ultralytics v8.3.123

Discussion
,
1

:rocket: Ultralytics v8.3.123 Release: INT8 RKNN Quantization, Improved Export, and More!

:glowing_star: Summary

We’re excited to announce Ultralytics v8.3.123! This release brings highly requested INT8 quantization support for Rockchip RKNN model exports, refined result reporting, upgraded TensorFlow compatibility, and a range of codebase and tracking improvements. Whether you’re developing on edge AI hardware, working on challenging datasets, or seeking smoother exports, this update has something for you. Check out the full Ultralytics v8.3.123 release notes for more details!

:triangular_flag: New Features

INT8 Quantization for Rockchip RKNN Exports

  • Deploy smaller, faster models: Export your YOLO models to INT8 for Rockchip devices, reducing model size and improving inference speeds—perfect for edge deployments!
  • Clearer model exports: Exported filenames now indicate whether you’ve chosen INT8 or FP16, making file management easier.
  • Try it out: Explore the new int8 export option in our documentation and experiment with optimized deployment on Rockchip platforms.

Thanks to @oDestroyeRo for adding RKNN INT8 quantization support!

:wrench: Improvements

  • Export Options Documentation: Export tables now document the new RKNN INT8 option for quick reference.
  • Function Renames for Clarity: Profiling-related methods have been renamed (profile → profile_ops, profile() → run()) for better clarity. See @glenn-jocher’s PR 20442.
  • Enhanced Results Output: The verbose() method on results now delivers more consistent, informative summaries for both detection and classification. See Optimize Result.verbose() by @glenn-jocher.
  • TensorFlow Export Fixes: Better compatibility and performance for TensorFlow SavedModel exports, especially supporting Attention blocks and GPU acceleration on Android. Special thanks to @Y-T-G for PR 20436.
  • Tracking Robustness: Improved feature extraction makes tracking more reliable—even when some images don’t contain detectable objects, thanks to @Y-T-G’s fix for IndexError.
  • Documentation and Usability: Many refinements for smoother developer experience and reduced confusion.

:hammer_and_wrench: Bug Fixes

  • Profiling Method Conflicts: Prevented conflicts caused by ambiguous method naming (PR 20442).
  • Empty ReID Indexing: Resolved IndexError when using feature extraction on frames with no detected objects (PR 20449).
  • TFLite GPU Delegate: Fixed performance issues with TFLite GPU delegate on Android (PR 20436).

:clap: New Contributors

A warm welcome to @oDestroyeRo for their first contribution to Ultralytics!

:inbox_tray: How to Upgrade and Get Involved

We encourage everyone to upgrade to v8.3.123, try out the new features, and let us know how it works for you. Your feedback drives our roadmap and helps the community thrive!

  • Read the full changelog: Compare v8.3.122…v8.3.123
  • Learn more about RKNN export and quantization: Check the updated documentation for instructions.

Questions, suggestions, or feedback? Just reply below or open an issue on Ultralytics GitHub Discussions.

Thanks for being a part of the YOLO community and making each release better!

— The Ultralytics Team