Ultralytics v8.3.160 Release Announcement!
Summary
We’re excited to announce Ultralytics v8.3.160! This release brings major improvements to keypoint handling, data augmentation, training workflows, exporting, and validation metrics. Whether you’re a developer customizing your pipeline or a researcher fine-tuning pose estimation, v8.3.160 delivers enhanced stability, accuracy, and usability for everyone working with YOLO models.
Explore the full release on GitHub Releases: v8.3.160.
Key Features & Improvements
Improved Keypoint Handling
-
Keypoint Clipping for Visualization
Keypoints are now clipped to remain within image boundaries, boosting training data quality and producing more reliable pose visualizations.
PR by @Laughing-q -
Original Keypoint Data Integrity
Keypoint data is preserved and confidence-filtering is more robust, leading to better training and prediction for pose/landmark tasks.
PR by @WillieMaddox -
Enhanced Keypoint Flipping
Vertical and horizontal flip augmentations for pose estimation now work more reliably, with clear user guidance if configs are missing.
PR by @WillieMaddox
Smarter Data Augmentation & Training
-
Intuitive Data Augmentation Access
Data augmentation transforms are now accessible and modifiable more seamlessly, making it much easier to customize and debug data pipelines. -
Improved Pretrained Weights Loading
Pretrained weights are now loaded directly within the training routine for a smoother and more predictable experience.
PR by @Laughing-q -
Parallel Training Compatibility
Text embedding generation (YOLOE & YOLO World) is now compatible with multi-GPU/Distributed Data Parallel setups, improving training stability.
PR by @Laughing-q
Export & Validation Enhancements
-
Dynamic Batch Export Fix
Exported models now handle dynamic batch sizes correctly, preventing shape mismatches and smoothing deployment flows.
PR by @Y-T-G -
Streamlined XML Export
Removed an unnecessary XML dependency for a lighter, more reliable experience.
PR by @Laughing-q -
Improved Validation Summaries
Summaries now show clearer metric names, per-class info, and include direct export options (such as Colab links), making model analysis easier.
PR by @RizwanMunawar
Other Notable Improvements
- YOLOE Predictor Optimization: Reduced redundant initialization for visual prompt predictions, boosting compatibility and efficiency.
PR by @Y-T-G - Consistent Object Counting & Documentation: Easier-to-understand counting results; enhanced tips for working with extreme aspect ratios in classification.
PR by @RizwanMunawar
PR by @picsalex
Bug Fixes
- Fixed edge cases in keypoint flipping and dynamic batch handling.
- Smoothed out multi-GPU training for YOLO World and YOLOE.
- Cleaned up object counting output and related documentation.
New Contributors
A warm welcome to @WillieMaddox for making their first contribution: preserving original keypoint data!
Try It Out & Give Us Feedback!
Ready to get started?
Update your Ultralytics package or clone the latest release and explore the new capabilities.
Your feedback drives this community forward—please share your thoughts, suggestions, and any successes (or hiccups) you encounter. Visit the Ultralytics Discussions forum or raise issues directly in the GitHub repository.
Thank you to the entire YOLO community and all contributors for making each release better. We look forward to seeing what you build next with YOLO and Ultralytics!
See the full changelog:
Compare v8.3.159…v8.3.160