New Release: Ultralytics v8.3.166

Discussion
, ,
1

:rocket: Ultralytics v8.3.166 Release: Dataset Overhaul & Evaluation Refinements!

:glowing_star: Summary

We’re excited to announce Ultralytics v8.3.166! This release brings streamlined dataset organization—especially for VisDrone—refinements in evaluation logic, and enhanced documentation for an even better user experience. Whether you’re training, validating, or benchmarking, these improvements help you work faster and smarter.

:bar_chart: New Features & Major Improvements

VisDrone Dataset Overhaul

  • Standardized Directory Structure:
    VisDrone data now downloads into easily navigable folders: images/train, images/val, and images/test.
    See PR by @glenn-jocher

  • Enhanced Annotation Conversion:
    The new conversion script reliably generates YOLO-format labels for each data split, reducing setup headaches.

  • Automated Cleanup:
    Original VisDrone folders are now automatically removed after conversion, giving you a tidy workspace.

Consistent Dataset Configurations

  • Unified Formatting:
    Comments and formatting across all dataset YAML files are now consistent, making configuration files easier to read and maintain for everyone.

  • Updated Download Size:
    The tiger-pose dataset now accurately reflects its actual download size: 49.8 MB.
    See PR by @Laughing-q

Evaluation Improvements

  • Oriented Bounding Box (OBB) Confusion Matrix
    Low-confidence detections are now excluded, and thresholding logic is refined, providing more reliable evaluation for OBB tasks.
    See PR by @Y-T-G

Documentation Enhancements

  • Contributor Information:
    Better author attribution and profiles in the documentation celebrates our contributors and makes it easier to follow their work.
    See PR by @glenn-jocher

:bullseye: Purpose & Impact

  • Cleaner Dataset Organization: The VisDrone update and cleanup tools reduce confusion and setup time for all users.
  • Consistency Across Datasets: New and experienced users alike can get up and running quickly with improved YAML files.
  • More Accurate Evaluations: The changes to OBB confusion matrix logic mean your advanced detection tasks yield more trustworthy results.
  • Transparent Download Sizes: No more surprises when working with datasets—planning storage is easier than ever.
  • Community Recognition: Improved documentation helps you connect with contributors and stay up-to-date with project changes.

:hammer_and_wrench: What’s Changed (PR Highlights)

Full changelog: Compare v8.3.165…v8.3.166
Full Release: Ultralytics v8.3.166 on GitHub

:speech_balloon: Try It Out & Share Feedback!

We invite you to test out v8.3.166 and let us know what you think! Your feedback helps guide future improvements and ensures Ultralytics continues to meet the needs of the entire YOLO community. If you run into any issues or have suggestions, please join the discussion here or open an issue on Ultralytics GitHub.

Thank you for being part of this journey—happy detecting!