New Release: Ultralytics v8.3.59

Discussion
,
1

:rocket: Announcing Ultralytics YOLO v8.3.59 Release!

We’re thrilled to unveil v8.3.59, the latest version of Ultralytics YOLO. This release brings groundbreaking features like TorchVision backbone integration and essential quality-of-life updates, including enhanced Docker support, dataset path refinements, and improved INT8 calibration validation. It’s all designed to empower developers with greater flexibility and smoother workflows. Let’s dive in! :star2:

:star2: Summary

What’s New in v8.3.59:

  • Custom TorchVision Backbone Support: Use any torchvision model (EfficientNet, MobileNet, etc.) as YOLO backbones with pretrained weights and layer customization.
  • Improved Segmentation Mask Support: .jpg mask compatibility added alongside .png support for seamless integration.
  • Docker Enhancements: JupyterLab setup made simpler and Docker image resilience improved with retry mechanisms.
  • Dataset Path Refinements: Organized YAML handling to reduce errors and improve usability.
  • Validation for INT8 Calibration: Robust checks enforce calibration dataset batch size requirements to prevent errors.
  • Enhanced Multi-Processing Documentation: Added troubleshooting for common Windows-related training errors.
  • New Performance Benchmarks: More metrics for NVIDIA DeepStream and Coral TPU development, including benchmarks for Raspberry Pi devices.

:bar_chart: Key Changes at a Glance

:fire: Features

  • TorchVision Integration: Seamlessly integrate popular TorchVision models like ResNet and ConvNext as YOLO backbones.
  • Expanded Workflow: .jpg mask support streamlines segmentation pipelines.

:hammer_and_wrench: Improvements

  • Docker Environment: Enhanced installation experience for Jupyter users, along with improved push stability.
  • Dataset Path Updates: Cleaner and more intuitive dataset YAML configuration.

:bug: Fixes

  • INT8 Export Validation: Calibration dataset checks ensure models are optimized without interruptions.
  • Documentation Updates: Added guidance for Windows multi-processing errors to facilitate trouble-free training.

:chart_with_upwards_trend: Benchmarks

  • NVIDIA Jetson DeepStream on Jetson devices.
  • Coral TPU benchmarks for Raspberry Pi 4B/5, extending edge hardware support.

:dart: A Closer Look at the Impact

Power and Flexibility :gear:

With the ability to use TorchVision models as backbones, developers can integrate state-of-the-art architectures alongside YOLO functionalities, accelerating object detection and classification tasks.

Streamlined Segmentation Pipelines :framed_picture:

The addition of .jpg mask support eliminates manual formatting delays, boosting efficiency in segmentation workflows.

Smarter Validation for Deployment :bulb:

Enhanced INT8 validation ensures hassle-free model compression and deployment, providing more robust setups for edge devices.

DevOps Simplification :whale:

Improved Docker tooling reduces friction for developers working across diverse environments.

:busts_in_silhouette: What’s Changed

:star2: New Contributors

:scroll: Full Changelog

Check out the complete v8.3.59 changelog for all the details.

We hope you enjoy this release! Feel free to try it out, explore the new TorchVision functionality, and give us your feedback. Your insights help us evolve and improve YOLO for everyone. :rocket:

:point_right: Download Ultralytics YOLO v8.3.59 here

Happy coding! :raised_hands:

2

This is an exciting update! :tada: The addition of TorchVision backbone integration in Ultralytics YOLO v8.3.59 is a game-changer, enabling seamless use of popular models like EfficientNet and ResNet with YOLO’s powerful features. This flexibility can significantly enhance custom object detection workflows by leveraging pre-trained weights and customizable layers.

The enhancements in segmentation mask support, including compatibility with .jpg masks, streamline workflows for segmentation tasks, reducing preprocessing overhead. Docker improvements and organized dataset path handling further simplify deployment and usage, especially for JupyterLab users.

For those focusing on edge devices, the inclusion of benchmarks for NVIDIA DeepStream and Coral TPU, along with validation enhancements for INT8 calibration, ensures efficient and robust deployment options.

If you’re exploring these new capabilities, don’t forget to update to the latest version and explore the improved documentation for multi-processing and training on Windows! :rocket: As always, feedback is welcome to help improve future updates. Happy experimenting! :blush: