New Release: Ultralytics v8.4.110

Discussion
, ,
1

:rocket: Ultralytics v8.4.110: RKNN Export for Every Supported Task

Ultralytics v8.4.110 expands RKNN export across all seven supported task categories and improves GPU-friendly tensor image plotting. This release makes Rockchip deployment more versatile while reducing unnecessary CPU transfers in accelerated visualization pipelines.

Explore the complete package details in the Ultralytics v8.4.110 release.

:glowing_star: New Features

RKNN export expanded to all supported tasks

RKNN export now supports models for:

  • Object detection
  • Instance segmentation
  • Classification
  • Pose estimation
  • Oriented bounding boxes (OBB)
  • Semantic segmentation
  • Depth estimation

The work from @glenn-jocher in PR #25490 also adds export smoke tests using rknn-toolkit2 across current task models.

Task-specific model names, export configurations, and generated artifacts are now documented more clearly, making deployment to RKNN-compatible Rockchip NPUs easier and more reliable. :mobile_phone:

:high_voltage: Performance Improvements

Keep tensor images on-device during mask plotting

Results.plot() and Annotator now accept contiguous HWC BGR torch.Tensor images.

With the improvements contributed by @glenn-jocher in PR #25487:

  • Tensor images can remain on their original device during mask compositing.
  • Unnecessary CPU transfers are avoided.
  • Conversion to NumPy is delayed until required by OpenCV, PIL, display, saving, or the returned result.
  • Existing NumPy image workflows remain fully supported.

This can improve visualization efficiency in GPU-accelerated segmentation pipelines without introducing major behavior changes. :artist_palette:

:books: Documentation and Fixes

This release includes several documentation updates:

  • Added task-specific RKNN export guidance for classification, depth, OBB, pose, segmentation, and semantic segmentation.
  • Updated plotting API documentation with tensor image support.
  • Corrected “setup” to “set up” in the DeepSparse tutorial through PR #25488 by @gizembm.

A warm welcome to @gizembm, who made their first contribution in this release! :tada:

:white_check_mark: Reliability

Expanded validation now covers RKNN conversion and basic inference across all seven supported task categories. These tests help ensure that exported models and their artifacts behave as expected across broader Rockchip deployment workflows.

:package: Try v8.4.110

Install or upgrade Ultralytics with:

pip install -U ultralytics

You can review every change between versions in the v8.4.109 to v8.4.110 full changelog.

Give the new RKNN export workflows and tensor plotting improvements a try, and let us know how they perform in your deployment and visualization pipelines! :raising_hands: