New Release: Ultralytics v8.3.134

[glenn-jocher]

Ultralytics v8.3.134 Release – Robust Tracking, Grayscale Support, and Easier Dataset Setup!

Summary

We’re excited to announce Ultralytics v8.3.134! This release focuses on key improvements for YOLO tracking reliability, grayscale image compatibility, better dataset path management, and model export stability. With cleaner logs, expanded testing, and polished solutions, this update offers a smoother and broader experience across Ultralytics tools.

---

New Features & Improvements

Tracking & Feature Extraction

• Reliable YOLO Tracking:
  No more duplicated feature hooks or memory leaks when using re-identification (ReID). Hooks now clean up as they should, resulting in stable, efficient tracking, especially with BoT-SORT.
  • PR: Fix ReID duplicated hooks when persist is False by @Y-T-G

Grayscale Image Support

• Better Grayscale Handling:
  Robust support for grayscale (single-channel) images in video streams, screenshots, tracking, and augmentations unlocks new workflows and camera/dataset compatibility.
  • PRs:
    Fix grayscale models tracking on videos by @Laughing-q
    Fix copy_paste for grayscale images by @Laughing-q

Dataset Path Resolution

• Simplified YAML Pathing:
  Dataset YAML files can now be referenced relative to the current directory, making configuration easier and reducing setup confusion.
  • PR: Search relative YAML dataset paths in training by @Y-T-G

Model Loading & Export

• Better Custom/Multi-Channel Model Loading:
  Loading pretrained weights is now more robust, especially for models using custom input channels.

  • PR: Fix pretrained weights loading for first convolution layer by @Laughing-q

• More Stable ONNX & RKNN Export:
  ONNX export is now restricted to versions below 1.18.0 to maintain compatibility (Pin onnx in pyproject.toml by @glenn-jocher), and INT8 quantization for Rockchip RKNN has been temporarily disabled (PR by @lakshanthad).

Solutions & Testing

• Multi-Person & Object Counting:
  Improved solutions for class-wise counting and AI Gym workout tracking, especially in crowded/multi-person environments. State tracking now uses defaultdict for accuracy and efficiency.

  • PRs:
    Use defaultdict for classwise object counting by @RizwanMunawar
    Replace list-based state tracking with defaultdict in AIGym by @RizwanMunawar

• Expanded Automated Testing:
  New and enhanced tests for result export formats and CLI solution modes to ensure ongoing reliability.

  • PRs:
    Add to_sql and to_html in pytests test_results method by @RizwanMunawar
    Add tests for solutions CLI examples by @RizwanMunawar

---

Other Improvements & Fixes

• Professional Log Messages:
  Cleaner output, including the removal of emojis for a more formal and consistent appearance.
  PR by @RizwanMunawar
• CI Enhancements:
  Conda jobs in continuous integration now have timeouts, preventing potential pipeline stalls.
  PR by @lakshanthad

---

Get the Details

• Full Changelog:
  See everything that changed in this release on the v8.3.134 GitHub compare page.
• Release URL:
  Ultralytics v8.3.134 Release

---

Get Involved & Share Feedback!

We’d love for you to try Ultralytics v8.3.134 and let us know how it works with your projects. Please share your feedback, questions, or suggestions in this forum—your insights help shape future releases and make the tools better for everyone.

Thank you to all community contributors and users for your support and energy!

— The Ultralytics Team