ultralytics 8.3.120

v8.3.120 - `ultralytics 8.3.120` CutMix augmentation fix via IoU overlaps (#20393)

on Python PyPI

🌟 Summary

This release brings smarter and more flexible data augmentation with a major CutMix upgrade, improved YOLOE training workflows, and enhanced support for all YOLO models in tracking and documentation. 🖼️🛠️

📊 Key Changes

• CutMix Augmentation Overhaul:
  • Added a num_areas parameter to CutMix, allowing multiple candidate regions for mixing images.
  • CutMix now avoids overlapping mixed regions with existing objects, reducing label noise and improving training data quality.
  • Improved logic for adding new objects during augmentation, ensuring better alignment and accuracy.
• YOLOE Model Improvements:
  • Unified and streamlined dataset handling and trainer inheritance for YOLOE and YOLO-World models, making training more consistent and maintainable.
  • Updated segmentation training and documentation to ensure correct trainer usage and robust loss calculations.
  • Standardized image channel handling for better dataset compatibility.
• Cleaner Augmentation Code:
  • Centralized the method for selecting random dataset indexes, removing duplicate code across augmentation classes.
  • Documentation for CutMix is now clearer and more complete.
• BOTSORT Tracker Update:
  • Tracker documentation and logic now refer to "YOLO" models in general, not just YOLOv8, clarifying support for all YOLO versions (including YOLO11, YOLO12, etc.).
  • Improved feature encoder logic for better handling of ReID (Re-Identification) settings.
• Workflow & Maintenance:
  • Upgraded CI dependencies for improved environment management.
  • .ts (TorchScript) files are now ignored in version control to prevent accidental commits.
  • New contributors recognized in documentation.

🎯 Purpose & Impact

• Better Model Training:
  • The smarter CutMix reduces label noise, leading to higher-quality training data and potentially more accurate, robust object detection models.
  • More control over augmentation with the new num_areas option benefits users seeking to fine-tune training behavior.
• Developer Experience:
  • Cleaner, more maintainable codebase makes it easier for developers to contribute and extend functionality.
  • Unified training logic for YOLOE and YOLO-World models simplifies custom training workflows.
• User Clarity:
  • Improved documentation and clearer naming help users understand features and capabilities, especially for new YOLO versions and tracking options.
• Broader Compatibility:
  • BOTSORT tracker and dataset handling updates ensure smooth support for all current and future YOLO models.
• Reliability:
  • Enhanced loss calculations and dataset initialization improve training stability, especially for segmentation tasks.

✨ This update is especially valuable for anyone training detection or segmentation models with Ultralytics, offering both immediate accuracy improvements and a smoother development experience!

What's Changed

• Bump astral-sh/setup-uv from 5 to 6 in /.github/workflows by @dependabot[bot] in #20390
• Remove redundant get_indexes functions in augment.py by @Laughing-q in #20389
• YOLOE: Fix visual prompt training by @Laughing-q in #20413
• YOLOE: Clean up duplicated modules by @Laughing-q in #20416
• Fix BOTSORT initialization with ReID set to False by @lwalejko in #20392
• ultralytics 8.3.120 CutMix augmentation fix via IoU overlaps by @Laughing-q in #20393

New Contributors

• @lwalejko made their first contribution in #20392

Full Changelog: v8.3.119...v8.3.120