New Release: Ultralytics v8.4.191

Discussion
, ,
1

:rocket: Ultralytics v8.4.191: Broader RKNN INT8 Support and Smarter Tracking

Ultralytics YOLO v8.4.191 is here! This release expands RKNN INT8 exports across vision tasks, reduces repeated work in segmentation training and parking management, improves tracking associations, and preserves prediction-source ordering.

Explore the v8.4.191 release notes for the full release details.

:glowing_star: Main Update: RKNN INT8 Export

PR #26904 by @Bovey0809 enables INT8 export for detection, segmentation, pose, and oriented bounding box (OBB) models.

The updated export splits inference work to address accuracy loss from INT8 box decoding:

  • NPU: Runs the head’s output convolutions.
  • CPU: Runs box decoding in floating point using a new tail.onnx file.
  • FP16-capable targets: Use RKNN’s auto_hybrid option to keep high-error layers in floating point.

:warning: RKNN Deployment Notes

  • Re-export older INT8 models to avoid incorrect coordinate scaling.
  • The decode tail requires ONNX Runtime, which Ultralytics installs automatically when needed.
  • Custom runtimes must run tail.onnx or implement equivalent decoding.
  • INT8 accuracy has been tested in the RKNN simulator, but has not yet been validated on Rockchip hardware.

:high_voltage: Performance Improvements

  • Faster segmentation training: PR #26914 by @raimbekovm finds foreground anchors once across the batch instead of repeating the work per image. Losses and gradients remain unchanged, with modest training-step improvements reported on a Tesla T4.
  • More efficient parking management: PR #26910 by @raimbekovm computes each tracked car’s center once per frame rather than once per parking-space comparison.

:hammer_and_wrench: Tracking and Prediction Fixes

  • More robust OC-SORT matching: PR #26912 by @mottopanikeiku measures detection direction over the track’s recent delta_t-frame window, making associations less sensitive to bounding-box jitter.
  • Prediction manifests keep your order: PR #26903 by @MohammadHijjawi97 preserves the listed order of images and videos in TXT and CSV sources instead of sorting them alphabetically—especially useful for frame sequences.

:books: Platform and Guide Updates

  • Annotation provider documentation: PR #26902 by @glenn-jocher adds MiniMax to the paid annotation providers documented for Ultralytics Platform.
  • Clearer import and label guidance: PR #26906 by @raimbekovm explains saved Roboflow and W&B import keys, along with OBB labels in detection datasets.
  • Updated training content: PR #26908 by @RizwanMunawar refreshes the Model Training Tips guide with an updated YOLO26 batch-processing video.

:rocket: Try It and Share Your Feedback

Upgrade to this release with:

pip install --upgrade ultralytics==8.4.191

Browse the full changelog to review all changes.

How does v8.4.191 work with your setup? Share your results, questions, or issues below—especially if you’re testing RKNN exports on Rockchip hardware. Thanks to all the contributors who helped make this release possible! :raising_hands: