ultralytics 8.3.160

v8.3.160 - `ultralytics 8.3.160` Clip `keypoints` for better visualization control (#21220)

on Python PyPI

🌟 Summary

This release brings improved keypoint handling, smarter data augmentation, and enhanced usability for training and exporting models, making YOLO workflows more robust and user-friendly. 🚀

📊 Key Changes

• Keypoint Clipping for Visualization: Keypoints are now clipped to stay within image boundaries, ensuring better training data quality and more accurate visualizations.
• Simplified Data Augmentation Access: Data augmentation transforms can now be accessed more intuitively, making it easier for developers to customize and debug pipelines.
• Improved Pretrained Weights Loading: Pretrained weights are now loaded directly during training, ensuring expected behavior and smoother user experience.
• Keypoint Data Integrity: The original keypoint data is preserved, and confidence-based filtering is handled more cleanly, improving data reliability.
• Enhanced Keypoint Flipping: Vertical and horizontal flip augmentations for pose estimation now work reliably, with clear warnings if required configuration is missing.
• Smarter Predictor Handling: YOLOE and visual prompt prediction logic is now more robust, reducing errors and improving compatibility with video and stream sources.
• Better Validation Metrics: Validation summaries now have clearer metric names, more detailed per-class info, and improved export options (including direct Colab links).
• Dynamic Batch Export Fix: Exported models now correctly handle dynamic batch sizes, preventing shape mismatches during inference.
• Parallel Training Compatibility: Text embedding generation is now compatible with multi-GPU setups, improving YOLO World and YOLOE training stability.
• Consistent Object Counting Results: Object counting results and documentation are now more consistent and easier to understand.
• Streamlined XML Export: The XML export process is simplified, removing unnecessary dependencies and ensuring more reliable output.
• Documentation Improvements: Expanded tips and examples for handling images with extreme aspect ratios in classification tasks.

🎯 Purpose & Impact

• More Reliable Keypoint Detection: By keeping keypoints within image bounds and preserving original data, users get higher-quality training and more accurate predictions, especially for pose estimation tasks.
• Easier Customization: Developers can now more easily access and modify data augmentations, leading to faster experimentation and fewer bugs.
• Smoother Training Experience: Loading pretrained weights and handling parallel training setups is now more intuitive and robust, reducing setup headaches.
• Better Results Analysis: Improved validation summaries and export options help both experts and newcomers quickly interpret and share model performance.
• Cleaner, Faster Exports: Streamlined XML export and dynamic batch fixes mean models are easier to deploy and integrate into diverse workflows.
• Clearer Documentation: Enhanced docs and code examples make it easier for all users to get started and avoid common pitfalls.

Overall, this update delivers a more stable, accurate, and user-friendly experience for anyone training, validating, or deploying YOLO models. 🎉

What's Changed

• Scope pretrained weights loading by @Laughing-q in #21156
• Preserve original keypoint data and eliminate internal 0.5 threshold by @WillieMaddox in #21165
• Support flip up/down augmentation for pose by @WillieMaddox in #20932
• YOLOE: Optimize redundant predictor initialization by @Y-T-G in #21198
• Update classification tip for CustomizedValidator by @picsalex in #21196
• Align metrics summary method to val logs by @RizwanMunawar in #21203
• Fix dynamic batch inference with nms=True by @Y-T-G in #21190
• YOLOE: Fix generate_text_embeddings for DDP training by @Laughing-q in #21210
• Add SolutionResults attributes in index.md and classwise_counts typo fix by @RizwanMunawar in #21216
• Remove unnecessary lxml dependency by @Laughing-q in #21221
• ultralytics 8.3.160 Clip keypoints for better visualization control by @Laughing-q in #21220

New Contributors

• @WillieMaddox made their first contribution in #21165

Full Changelog: v8.3.159...v8.3.160