Ultralytics v8.3.25 Release Announcement
Summary
We’re thrilled to announce the release of Ultralytics v8.3.25, packed with exciting new features and improvements. This update introduces support for the Alibaba MNN (Mobile Neural Network) format, enhancing the deployment of YOLO models on mobile and embedded systems.
Key Changes
New Features
- Alibaba MNN Support: You can now export and predict with YOLO models in the MNN format, making mobile deployments more efficient.
Improvements
- ONNX Runtime Optimization: Experience faster inference with reduced overheads, supporting both dynamic and static shape inferences.
- Tracking Enhancements: Default confidence thresholds for trackers have been lowered to better align with detection predictions.
Bug Fixes
- TFLite Image Size Fix: Arbitrary image sizes for TFLite are now supported, thanks to @ambitious-octopus.
- ONNX Runtime Example Update: Updated to support
onnxruntime==1.19.xby @yawnBright.
Purpose & Impact
- Enhanced Mobile Deployment: MNN support allows for high-efficiency model deployment on mobile and ARM devices.
- Performance Optimization: Faster ONNX inference is crucial for real-time applications.
- User-Friendly Tracking: Updated thresholds make tracking models more intuitive and aligned with user expectations.
What’s Changed
- Fix arbitrary imgsz for TFLite by @ambitious-octopus
- Example ORT==2.0.0-rs.5 to support onnxruntime==1.19.x by @yawnBright
- Update Triton Inference Server guide by @Y-T-G
- Faster ONNX inference with bindings by @Y-T-G
- Decrease default confidence threshold for tracking by @Y-T-G
- Alibaba MNN export and predict support by @wangzhaode
New Contributors
- Welcome @yawnBright and @wangzhaode for their first contributions!
We encourage you to try out the new release and share your feedback. Your insights are invaluable in helping us improve and innovate.
Full Changelog: v8.3.25 Changelog
Explore the new features and improvements by visiting the release page. Happy experimenting!