Ultralytics v8.3.159 Release – Unified COCO Evaluation, Jetson Enhancements, and More!
Summary
We’re excited to announce the release of Ultralytics v8.3.159, packed with significant improvements aimed at streamlining the codebase and enhancing the user experience. This update unifies the COCO evaluation logic for detection, segmentation, and pose tasks, brings targeted improvements for edge devices like NVIDIA Jetson, and introduces cleaner analytics, better dependency management, and refined documentation. Whether you’re developing, benchmarking, or deploying YOLO models, there’s something here for you!
New Features & Major Improvements
Unified COCO Evaluation Logic
Detection, segmentation, and pose validators now share a single coco_evaluate method. This centralization reduces code duplication, simplifies future maintenance, and ensures all metric reporting is consistent and reliable.
See the PR by @Laughing-q
Object Counter Display Fix
The object counter analytics are now clearer—“IN” and “OUT” counts are only displayed when their respective options are enabled. This makes interpreting your results easier than ever.
See the update by @fn-hide
Similarity Search Refactor
The underlying CLIP text model now supports direct image encoding and adopts a unified model-loading approach, making similarity search more modular, easy to maintain, and extensible.
See the PR by @RizwanMunawar
Improved Experiment Tracking
Validation metrics for detection models now include the save_dir parameter, making tracking experiments and managing your results more straightforward.
Details in @RizwanMunawar’s PR
Other Notable Changes
NVIDIA Jetson Documentation Update
Jetson YOLO11 benchmarks have been refreshed using the larger COCO128 dataset, and support for the MNN model format has been added. This ensures more reliable, representative benchmarks for edge deployments.
Update by @lakshanthad
Dependency Pinning for IMX
The Model Compression Toolkit is now pinned to versions >=2.3.0 and <2.4.1, offering greater stability for IMX export and ONNX IMX inference, especially for Sony IMX hardware users.
See the update by @Laughing-q
Documentation & Link Updates
We’ve cleaned up and updated numerous documentation links (including tracker and language menu links), and improved 301 redirects for more accurate and user-friendly navigation.
Update by @glenn-jocher
Full Changelog & Contributors
- Fix count display check for
show_in=Falseandshow_out=Falseby @fn-hide - Add
save_dirtoMetricsfor better access by @RizwanMunawar - Update NVIDIA Jetson Doc with COCO128 Benchmarks by @lakshanthad
- Update 301 redirects by @glenn-jocher
- Use
TextModelclass for similarity search by @RizwanMunawar - Pin Model Compression Toolkit for IMX export by @Laughing-q
- Refactor and Clean up COCO evaluation by @Laughing-q
New Contributor:
Welcome and thank you to @fn-hide for their first contribution!
Full Changelog:
Compare v8.3.158…v8.3.159 on GitHub
How to Get Started
Install or upgrade Ultralytics via pip:
pip install -U ultralytics
Explore the v8.3.159 Release Notes for more detail.
Get Involved — We Value Your Feedback!
We encourage all users to try out this latest release, especially if you’re working across multiple YOLO tasks or benchmarking on Jetson and IMX platforms. Please report your experiences, suggestions, or any issues in our GitHub Discussions or by opening an Issue.
Your feedback plays a pivotal role in driving YOLO forward—thank you to the whole community and the amazing Ultralytics team!
Happy experimenting!