Ultralytics v8.3.157 Release: Lightning-Fast Evaluation & Enhanced Usability!
Summary
We’re excited to announce Ultralytics v8.3.157—a major update focused on making your model development experience faster, easier, and more flexible! This release brings dramatically improved evaluation speeds, smarter segmentation & pose handling, greater dataset compatibility, and even better documentation to accelerate your workflows. Whether you’re a newcomer or a seasoned YOLO developer, you’ll notice the difference right away.
New Features & Major Improvements
Major Evaluation Upgrade
- Faster COCO/LVIS Evaluation: We’ve integrated the
faster-coco-evallibrary by @MiXaiLL76 as a drop-in replacement forpycocotools. Enjoy up to 4.5x faster validation on COCO and LVIS datasets! Benchmarking and model iteration just got a huge boost.
Segmentation & Pose Enhancements
- Mask Post-Processing: No more custom mask headaches! Changes by @Laughing-q fix mask shape mismatches, making YOLO segmentation more robust—especially with custom pipeline modifications.
- Reliable Keypoints: Pose models now treat keypoint labels and predictions in a consistent format, reducing errors and making training and inference more dependable.
Dataset Flexibility
- Grounding Dataset Improvements: Thanks to @mohiuddin-khan-shiam, label verification is now smarter: known datasets are strictly validated, but custom datasets load more gracefully. No more loading interruptions!
Documentation & UX Improvements
- Video Tutorial for Beginners: A new dog pose estimation tutorial added by @RizwanMunawar is now featured in the docs—perfect for quick onboarding and visual learners!
- Clarity in Dataset Guides: @glenn-jocher updated GQA dataset links for YOLOE and YOLO-World, ensuring you always have the latest references.
- Easier Data Augmentation: Solutions docs now include a direct button linking to the YOLO augmentation guide (thanks @RizwanMunawar).
Dependency Updates & Fixes
-
Security & Compatibility:
- Flask is now pinned to version 3.0.1+ for similarity search (PR by @RizwanMunawar).
- urllib3 upgraded to 2.5.0 in RTDETR ONNXRuntime examples.
- OpenVINO export for macOS requires version 2025.2.0+ for full compatibility (PR by @glenn-jocher).
-
Better Error Handling:
- Validation now catches missing label/prediction files more elegantly (PR by @Y-T-G).
- Removed unnecessary conditional checks for cleaner code (PR by @WillieMaddox).
New Contributors
Big thanks to our newest contributors for their valuable additions:
You’re fueling innovation for the whole YOLO community!
How to Get Started
Update now to the latest version (pip install -U ultralytics) and experience faster evaluation, improved segmentation and pose workflows, and slicker documentation.
We encourage all users to review the full changelog, try out the new features, and let us know about your experience or suggestions right here in the forum.
Feedback Welcome
Your input drives our roadmap! Please share your thoughts, issues, or success stories—every bit helps us build better tools for you and the whole computer vision community.
Check out the full release announcement: Ultralytics v8.3.157 Release Notes
Happy training!
— The Ultralytics Team & YOLO Community