New Release: Ultralytics v8.3.120

Discussion
,
1

:rocket: Ultralytics v8.3.120 Release — Smarter CutMix, YOLOE Training Upgrades, and More!

:glowing_star: Summary

We’re excited to announce Ultralytics v8.3.120! This release brings a major upgrade to CutMix augmentation, improved YOLOE model training workflows, broader and clearer tracking/documentation support for all YOLO models, and a cleaner, more maintainable codebase. Whether you’re a new user or a power developer, you’ll enjoy better training accuracy, improved flexibility, and smoother workflows.

:new_button: New Features & Improvements

CutMix Augmentation Overhaul

  • Multiple Mixed Regions: CutMix now supports a num_areas parameter, letting you mix several regions per image—great for advanced augmentation strategies.
  • Smarter Object Handling: Overlapping of mixed regions with existing objects is avoided, reducing label noise and ensuring higher training data quality.
  • Enhanced Logic: Objects are now added more accurately during augmentation, with better alignment for improved performance.
  • Cleaner Code: Centralized dataset index selection removes duplication and streamlines all augmentation routines.
  • Improved Documentation: CutMix docs now provide clearer guidance and usage examples.

YOLOE and YOLO-World Improvements

  • Unified Training Workflows: Dataset handling and trainer inheritance for YOLOE and YOLO-World are now streamlined, making training more consistent and maintainable.
  • Segmentation Upgrades: Updated training and docs ensure robust segmentation loss calculations and correct trainer usage.
  • Standardized Channel Handling: Improved image channel support for broader dataset compatibility.

Broader & Clearer Tracking Support

  • BOTSORT Tracker Update: Tracker logic and documentation now consistently refer to “YOLO” models, strengthening support for all YOLO versions (YOLO11, YOLO12, more).
  • Feature Encoder Logic Polished: Better handling of ReID (Re-Identification) settings in tracking.

Maintenance & Workflow Enhancements

  • Upgraded CI Dependencies: Improved environment management for greater reliability.
  • .ts File Ignore: TorchScript files are now excluded from version control.
  • New Contributors: Welcome @lwalejko, who made their first contribution in this release!

:bullseye: Purpose & Impact

  • Better Model Training: Reduced label noise and higher-quality augmentations mean more robust object detection and segmentation results.
  • Greater Control & Transparency: Augmentation and workflow controls are more flexible; improved docs and naming clarify functionality across all models.
  • Enhanced Developer Experience: Cleaner, unified codebase makes it easier for you and the community to extend and contribute.
  • Seamless Multi-Model Support: Tracker and dataset changes pave the way for smooth integration of all current and future YOLO model versions.
  • Reliability: Segmentation and dataset handling are more stable—train with confidence!

:hammer_and_wrench: What’s Changed

Full Changelog: Compare v8.3.119…v8.3.120

:clap: Thank You, New Contributors!

A special shoutout to @lwalejko for joining the Ultralytics open-source community with their first PR!

:test_tube: Try It Out and Share Your Feedback!

We encourage everyone to download Ultralytics v8.3.120 from the official release page, explore the new features, and let us know your thoughts right here. Your suggestions, bug reports, and ideas help drive YOLO innovation for everyone!

On behalf of the entire Ultralytics team and the broader YOLO community—thank you for your support and collaboration!

Happy training! :rocket: