Ultralytics v8.3.167 Release – Expanded Sony IMX Support, Smoother Visual Prompting, and More!
Summary
We’re excited to announce Ultralytics v8.3.167, a release packed with valuable improvements. This version enhances Sony IMX device compatibility (now with pose estimation!), streamlines visual prompting workflows, and includes documentation and usability refinements—making edge AI and model deployment even more accessible. Dive into the details below!
New Features & Improvements
Sony IMX Export Enhancements
- Detection & Pose Estimation Support:
You can now export both object detection and pose estimation models for Sony IMX devices, widening the range of edge AI applications. - Required Dependency:
Addedmct-quantizers>=1.6.0
to ensure IMX export reliability. - Improved Post-processing:
Pose estimation outputs are now handled correctly during both export and inference on IMX devices. - Documentation:
The Sony IMX500 integration guide now includes pose examples, clearer setup, and updated visuals.
Visual Prompting Workflow Smoothing
- Persistent Class Selections:
In YOLOE, classes selected using visual prompts for image references now persist through future predictions and exports, saving time and effort.
Learn more in the PR.
Confusion Matrix Plot Optimization
- Scalability Fixes:
Confusion matrix plotting now robustly handles 100+ classes—keeping visuals readable and metrics accurate.
Thanks to @Laughing-q (PR).
Documentation & Workflow Boosts
- Clarity & Coverage:
- Expanded Sony IMX docs and refined IBM watsonx integration guidance
- YOLOE visual prompt behavior now transparently detailed
- Reliability Upgrades:
- Upgraded Slack GitHub Action (PR) to v2.1.1 for better notifications
- Version bumped to 8.3.167
Purpose & Impact
- Broader Edge AI Reach:
With both detection and pose models supported on Sony IMX, deploying advanced models at the edge just got easier. - Easier Usability:
Persisting visual prompt classes and clearer docs make workflows faster and more intuitive—especially for new or non-technical users. - Analysis at Scale:
Confusion matrix plots remain practical and insightful even with large class counts. - Stable DevOps:
Workflow reliability improvements mean less risk of export or notification hiccups.
What’s Changed (Pull Requests)
- Fix
self.names
overwriting inConfusionMatrix
by @Laughing-q - YOLOE: Document visual prompt persistence for prediction with refer_image by @Y-T-G
- Bump slackapi/slack-github-action from 2.1.0 to 2.1.1 by @dependabot[bot]
- Support IMX export and inference for Pose Estimation by @ambitious-octopus
- Fix Sony IMX export
mct-quantizers>=1.6.0
dependency by @ambitious-octopus
Full changelog: 8.3.166…8.3.167
Release notes: v8.3.167 release page
Try It Out & Share Feedback!
Upgrade to Ultralytics v8.3.167 and explore the new features—whether you’re deploying to edge IMX hardware, exploring pose estimation, or streamlining workloads with visual prompts. Your feedback helps the YOLO community and Ultralytics team keep making these tools better for everyone!
Thanks for being part of the journey—happy building!