ultralytics 8.3.132

v8.3.132 - `ultralytics 8.3.132` Always transfer `Conv` layer pretrained weights (#20567)

on Python PyPI

🌟 Summary

This release brings smarter model weight loading for multi-channel training, introduces the HomeObjects-3K indoor dataset, and enhances object counting, segmentation, and documentation for a more robust and user-friendly Ultralytics experience. 🚀🏠

📊 Key Changes

• Smarter Model Weight Loading:

  • The model now intelligently transfers pretrained weights for the first convolutional layer, even when input channel sizes differ. This means easier adaptation to datasets with different image formats (e.g., multi-channel images).
  • Improved logging gives clearer feedback on which weights were successfully loaded.

• New HomeObjects-3K Dataset:

  • Added a high-quality indoor object detection dataset with 12 common household items (like beds, sofas, TVs, and more), ideal for smart home, robotics, and AR applications.

• Object Counting for Rotated Boxes (OBB):

  • The object counting solution now supports rotated bounding boxes, improving tracking and counting accuracy for objects at various angles—especially useful for aerial or industrial imagery.

• Improved Segmentation Mask Handling:

  • Instance segmentation workflows now reliably extract and provide masks, reducing errors and making segmentation tasks more robust.

• Unified Dataset Handling & Validation:

  • Training pipelines now use a consistent data structure across all YOLO tasks, with better error handling for pose datasets.

• Branding & Documentation Updates:

  • All references updated from "YOLOv8" to "Ultralytics YOLO" for consistent branding.
  • A new YouTube video tutorial on data preprocessing and augmentation is now embedded in the docs for easier onboarding.

• Streamlined CI & Maintenance:

  • YOLOv10 benchmarks removed from Raspberry Pi CI to focus on newer models.
  • Upgraded Slack notification integration for CI workflows.

🎯 Purpose & Impact

• Greater Flexibility:

  • Users can now seamlessly use pretrained models with datasets that have different image channel counts, reducing manual work and errors.

• Expanded Dataset Choices:

  • The HomeObjects-3K dataset makes it easier to train and evaluate models for indoor environments, supporting a wide range of real-world applications.

• Improved Tracking & Counting:

  • Rotated bounding box support ensures more accurate object tracking and counting, especially in complex scenes.

• Better User Experience:

  • Enhanced documentation, clearer logs, and video tutorials make it easier for both beginners and experts to get started and troubleshoot.

• Consistency & Professionalism:

  • Unified branding and codebase improvements ensure a polished, reliable experience for all users.

✨ This update is packed with improvements that make Ultralytics YOLO models more adaptable, user-friendly, and ready for advanced computer vision tasks—whether you're working in research, industry, or just getting started!

What's Changed

• New HomeObjects-3K dataset by @RizwanMunawar in #20574
• Add https://youtu.be/IYWQZvtOy_Q to docs by @RizwanMunawar in #20562
• Update YOLOv8 references to Ultralytics YOLO in docstrings by @RizwanMunawar in #20563
• Remove YOLOv10 from Raspberry Pi Benchmarks CI by @lakshanthad in #20582
• Scope getting masks from segmentation model by @RizwanMunawar in #20584
• Bump slackapi/slack-github-action from 2.0.0 to 2.1.0 in /.github/workflows by @dependabot[bot] in #20591
• Add object counting support for OBB task by @RizwanMunawar in #20585
• Add kpt_shape validation in Pose Trainer with clear error handling by @RizwanMunawar in #20547
• ultralytics 8.3.132 Always transfer Conv layer pretrained weights by @Laughing-q in #20567

Full Changelog: v8.3.131...v8.3.132