Ultralytics v8.3.132 Release: Smarter Weight Loading, New Indoor Dataset, and More!
Summary
Ultralytics v8.3.132 is here! This release brings smarter pretrained model handling for multi-channel images, a brand new HomeObjects-3K indoor dataset, enhanced object counting (now with rotated boxes), and numerous improvements to segmentation and documentation. Whether you’re developing in smart home, robotics, or cutting-edge computer vision, this update makes Ultralytics YOLO even more robust and user-friendly!
New Features
Smarter Model Weight Loading
- Pretrained weights for the first convolutional layer now transfer intelligently, even if your dataset’s input channel size differs.
- Clearer logging informs you which weights were successfully loaded for maximum transparency.
- Always transfer Conv layer pretrained weights by @Laughing-q
HomeObjects-3K Indoor Dataset
- Introducing a new high-quality dataset with 12 common household objects—such as beds, sofas, and TVs—ideal for indoor detection in smart home, AR, and robotics projects.
- New HomeObjects-3K dataset by @RizwanMunawar
Object Counting for Rotated Boxes (OBB)
- Object counting now supports rotated bounding boxes, greatly enhancing accuracy for aerial, industrial, and non-axis-aligned scenes.
- Add object counting support for OBB task by @RizwanMunawar
Enhanced Segmentation Mask Workflows
- Segmentation models now provide instance masks more reliably, resulting in more robust downstream segmentation tasks.
- Scope getting masks from segmentation model by @RizwanMunawar
Improvements & Fixes
-
Unified Dataset Handling & Validation: Training now uses a consistent dataset structure for all tasks, with improved error handling for pose task validation.
-
Docs, Branding, and Onboarding: References to “YOLOv8” updated across documentation for unified branding. A new YouTube tutorial on data preprocessing and augmentation makes onboarding smoother.
-
CI/CD & Maintenance:
- YOLOv10 benchmarks have been removed from Raspberry Pi workflows to streamline CI for newer models (Remove YOLOv10 from CI by @lakshanthad).
- Slack integration for CI is upgraded for better team notifications (Slack action bump by @dependabot[bot]).
Purpose & Impact
- Adapt Models to Any Dataset: Pretrained weights now integrate more smoothly with custom datasets, no matter the image channel configuration.
- Expand to Indoor Applications: HomeObjects-3K dataset opens new possibilities for smart, AR, and robotics applications in households and indoor environments.
- Precision with Rotated Objects: Reliable tracking and counting in industrial, aerial, and off-axis scenes with OBB support.
- Better User Experience: Improved error messages, onboarding videos, and detailed logs smooth the path for both new and advanced users.
- Professional and Consistent Branding: “Ultralytics YOLO” branding and cleanup across the board enhance project clarity.
Try it Out!
You can install or upgrade to v8.3.132 today:
pip install -U ultralytics
Try out the new features, explore the HomeObjects-3K dataset, or adapt a pretrained model to your own multi-channel data. We hope these changes make your workflow smoother and your models even stronger!
Feedback Wanted!
Your feedback drives Ultralytics forward! Please share your experiences, feature requests, and questions right here in Discourse. Reporting issues or successes helps the community and the Ultralytics team improve with every release.
For a deep dive, see the full changelog and detailed GitHub release notes.
Thanks for being part of this journey—the progress is a collective achievement from the entire YOLO community and Ultralytics team!
Stay curious and keep building!