New Release: Ultralytics v8.3.72

Discussion
,
1

:rocket: Ultralytics Release v8.3.72 Announcement

We are excited to announce the release of Ultralytics v8.3.72, packed with enhancements for NVIDIA Jetson DLA inference, improved export functionalities, and a host of optimizations and fixes. This release enhances usability, performance, and documentation to ensure a smoother experience for all users. Here’s a closer look at what’s new:

:star2: Summary

The v8.3.72 release focuses on:

  • Enhancing NVIDIA Jetson DLA support for improved inference compatibility.
  • Overhauling export documentation for better clarity and configurability.
  • Optimizing segmentation rendering for improved label handling.
  • Fixing bugs to improve reliability, particularly for advanced setups like multi-GPU training.
  • Providing enriched resources for the Crack Segmentation Dataset.

:bar_chart: Key Changes

New Features

  1. Enhanced NVIDIA Jetson DLA Support:

    • Added explicit control for DLA core selection (e.g., dla:0/dla:1) during TensorRT export and inference.
    • Improved documentation for NVIDIA Jetson DLA, detailing device specs like core count and frequency.
    • Fixed issues with metadata for DLA-specific inference for smoother setups on Jetson devices.

  2. Export Documentation Overhaul:

    • Added comprehensive argument tables to export docs for formats like ONNX, TensorRT, and CoreML.
    • Now includes explanations for configuration options like FP16, INT8 quantization, and dynamic input sizes, enabling easier optimization for diverse hardware.

Improvements

  • Optimized seg_bbox Rendering:

    • Improved label handling in the plotting utility for segmentation tasks, reducing unnecessary computations. Slight performance gains can be observed during visualization.

  • Streamlined Crack Segmentation Documentation:

    • Added Colab integration, a tutorial notebook, and a demo video for the Crack Segmentation Dataset, making it more accessible to infrastructure and safety researchers.

Bug Fixes

  1. Fixed a missing nc attribute error during NMS export, improving reliability for custom models and multi-GPU setups.
  2. Improved dependencies by replacing beautifulsoup4 with mkdocs-ultralytics-plugin>=0.1.17 to streamline documentation builds.
  3. Addressed warnings and inefficiencies for better code stability.

:dart: Purpose & Impact

Why Upgrade?

  • Improved Compatibility: NVIDIA Jetson DLA updates improve edge AI device compatibility for more seamless inference setups.
  • Simplified Exports: Enhanced export docs clarify configurations, empowering users to tailor models easily for any deployment scenario.
  • Faster Visualization: Optimizations reduce runtime overhead, improving segmentation and plotting speed.
  • Enhanced Reliability: Fixed issues ensure a smoother experience, especially for complex systems like multi-GPU or custom-designed models.
  • Easier Onboarding: Comprehensive Crack Segmentation resources help researchers get started quickly with actionable datasets.

:raised_hands: Acknowledgments

Special thanks to our community contributors for their stellar efforts in this release:

Full list of changes: v8.3.72 Changelog

:star2: How to Get Started

Update your Ultralytics package to the latest version:

pip install ultralytics --upgrade

:speech_balloon: Join the Conversation

We would love to hear your feedback! Try out the latest release and let us know your thoughts, experiences, and suggestions in the Ultralytics Discussions.

Thank you for being part of the YOLO and Ultralytics journey. Your contributions and collaboration fuel every improvement! :rocket: