🌟 Summary

The v8.3.25 release of Ultralytics brings a major update with the introduction of support for the Alibaba MNN (Mobile Neural Network) format, enabling better deployment of YOLO models on mobile and embedded systems.

📊 Key Changes

• Alibaba MNN Support: Added the ability to export and predict with YOLO models in the MNN format.
• Integration Updates: Documentation and integration guides have been enhanced to include MNN details.
• Improved ONNX Runtime: The ONNX Runtime has been optimized for faster inference, reducing overheads and improving performance, allowing for efficient handling of dynamic and static shape inferences.
• Tracking Improvements: Lowered default confidence thresholds for trackers to align with detection predictions, facilitating better tracking performance.

🎯 Purpose & Impact

• Enhanced Mobile Deployment: With MNN support, users can deploy models with high efficiency on mobile and ARM devices, allowing for more versatile applications.
• Performance Optimization: Faster ONNX inference ensures reduced runtime and improved speed, which is particularly beneficial for applications requiring real-time processing.
• User-Friendly: Updated thresholds for tracking models now provide more intuitive operations, aligning the outputs of tracking systems with user expectations.
• Streamlined Updates: Improved configuration and integration steps reduce friction during the setup of modern AI workflows, enhancing overall usability and user satisfaction.

What's Changed

• Fix arbitrary imgsz for TFLite by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/17138
• Example ORT==2.0.0-rs.5 to support onnxruntime==1.19.x by @yawnBright in https://github.com/ultralytics/ultralytics/pull/16962
• Update Triton Inference Server guide by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/17059
• Faster ONNX inference with bindings by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/17184
• Notify only on first CI run by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17241
• Decrease default confidence threshold to start tracking new tracks by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/17172
• Update publish.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17251
• Pin ray numpy<=2.0.0 test by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/17245
• Update notebooks by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17260
• Fix missing argument by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/17253
• Update triton-inference-server.md by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/17252
• Disable Ray tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17266
• ultralytics 8.3.25 Alibaba MNN export and predict support by @wangzhaode in https://github.com/ultralytics/ultralytics/pull/16802

New Contributors

• @yawnBright made their first contribution in https://github.com/ultralytics/ultralytics/pull/16962
• @wangzhaode made their first contribution in https://github.com/ultralytics/ultralytics/pull/16802

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.24...v8.3.25