New Release: Ultralytics v8.3.164

[glenn-jocher]

Ultralytics v8.3.164 Release Announcement

Summary

We’re excited to announce the release of Ultralytics v8.3.164! This update brings critical improvements to YOLO detection validation, increased dataset flexibility, smoother export experiences, and refined documentation and developer usability—all aimed at making your computer vision workflows faster, more accurate, and easier to use.

This release is recommended for all users, especially those validating detection models, working with diverse datasets, or deploying with advanced export requirements.

---

New Features & Major Improvements

Metrics Fix in YOLO Detection Validation

• A long-standing issue with swapped mAP50 and mAP50-95 metrics in detection validation is now fixed, ensuring your validation results accurately reflect your model’s performance!
  • PR by @Y-T-G: Fix swapped mAP50 and mAP50-95 in COCOEval stats

Flexible Text Sample Handling in Datasets

• You can now control the number of text samples in GroundingDataset using the max_samples parameter, allowing finer dataset customization and improved negative text sample selection logic.
  • PR by @Y-T-G: Fix inconsistent max_samples value

Export Improvements & Warnings

• TensorRT export is safer and more reliable, with better CUDA device handling and explicit warnings for dynamic batch sizes—helping prevent common pitfalls during deployment.
  • PR by @Y-T-G: Add max batch warning for dynamic model with NMS

Classification Dataset Compatibility

• The framework now recognizes valid/ as a fallback validation folder for classification tasks, improving compatibility with Roboflow and similar dataset exports.
  • PR by @JamesBond6873: Support valid/ as fallback folder

HEIC Image Support Update

• Switched from pillow_heif to pi-heif for decoding HEIC images, simplifying license requirements for seamless integration in open source and commercial projects.
  • PR by @picsalex: Switch pillow_heif to pi-heif

---

Additional Improvements

• Type Hints: Standardized the use of np.ndarray, making type annotations clearer and more consistent throughout the codebase (PR by @Laughing-q).
• Documentation & UI Enhancements:
  • Embedded a new YouTube video guide into the data annotation documentation (PR by @RizwanMunawar).
  • Improved example for the sweep_annotator method (PR by @RizwanMunawar).
  • Cleaned up CSS from the Similarity Search web page for a snappier interface (PR by @RizwanMunawar).
• Minor Fixes: Corrected a misleading error message in classification augmentation for a smoother developer experience (PR by @Toprak2).

---

New Contributors

A big thank you to our new contributors for their valuable efforts:

• @JamesBond6873 for improving classification dataset folder support
• @Toprak2 for clarity fixes in augmentations

Your feedback, testing, and pull requests truly power the evolution of Ultralytics. Kudos to everyone in the community—the real strength of YOLO!

---

How to Upgrade

Upgrade as usual using pip:

pip install --upgrade ultralytics

View the full release details at the v8.3.164 release page.
See all changes in the changelog.

---

Get Involved & Share Feedback

We encourage you to try out v8.3.164, especially if you’re working with custom datasets, validating models, or exporting for deployment. Please share your experiences, suggestions, or questions—the community and Ultralytics team are always ready to help!

Happy detecting!