v2.6.0#

v2.6.0 adds an Edge Impulse uploader for sending labeled datasets straight from the Dataset Editor, brings up Arduino Portenta H7 board support, and migrates the bundled machine-learning workflow from the legacy CMSIS-NN models to TensorFlow Lite. This release stays on the Qt Creator 4.0.2 base used by the prior 2.x releases. It contains one user-visible workflow break: the old CMSIS-NN .network models and their nn_* examples have been removed and the MobileNet models have moved.

Highlights#

  • Edge Impulse uploader lets you log into your Edge Impulse account, pick a project, choose a train/test split, and upload the open dataset directly from the Dataset Editor.

  • Arduino Portenta H7 (STM32H747) is now a supported board with bundled firmware, bootloader, and UVC images.

  • TensorFlow Lite migration replaces the legacy CMSIS-NN neural-network models and nn_* examples with the TensorFlow Lite workflow.

  • New ML examples add TensorFlow Lite face data collection and on-device face recognition.

  • Bundled firmware updated from 3.6.4 to 3.6.7 for all boards.

New features#

  • Edge Impulse uploader. Added an uploader that logs into your Edge Impulse account, lets you pick a project and a train/test split, and uploads the open dataset directly from the Dataset Editor (or by API key) as signed JPEG ingestion samples. The JWT/API key is stored in QSettings (8a828b45b, fa86179a7, ca9183864, a267de659).

  • TensorFlow Lite face examples. Added tf_face_collection.py and tf_face_recognition.py under 25-Machine-Learning for face data collection and on-device face recognition/classification (fb2d2cbb9).

  • Readout-Control example. Added an apriltag_tracking.py example under 35-Readout-Control demonstrating high-speed AprilTag tracking by moving a small sensor readout window on the OpenMV Cam H7 Plus (OV5640) (0c6ce1a82).

Other changes and improvements#

  • Replaced the legacy CMSIS-NN neural-network models and nn_* examples with the TensorFlow Lite workflow, and relocated the bundled MobileNet models from share/qtcreator/models/tensorflow/mobilenet to share/qtcreator/models/mobilenet (0c6ce1a82).

Bug fixes#

  • Fixed the Dataset Editor holding a lock on a closed dataset folder by rebuilding the file-system model when the folder is closed, and delayed window-state restore so the file-system watcher initializes correctly (af9fb441b).

  • Fixed the dataset export menu so the single Export Dataset to Zip File action works correctly by removing the dead duplicate export entry, and corrected error-dialog/progress ordering on export failure (ca9183864).

  • Fixed the Linux .png application icon not being produced in the release build (a55dd4c1b).

Platform and tool support#

  • Qt Creator base: 4.0.2.

  • New board: Arduino Portenta H7 (STM32H747), with bundled firmware/bootloader/UVC images (54f4d565b).

  • Bundled OpenMV firmware: updated from 3.6.4 to 3.6.7 for OPENMV2/M4, OPENMV3/M7, OPENMV4/H7, OPENMV4P/H7 Plus, and PORTENTA (54f4d565b, fb2d2cbb9, 4bb1cf8d5).

Breaking changes#

The bundled CMSIS-NN/legacy neural-network models (cifar10, cifar10_fast, lenet, smile, chars74k .network files) and their nn_* example scripts have been removed in favor of the TensorFlow Lite workflow. The MobileNet models have moved from share/qtcreator/models/tensorflow/mobilenet to share/qtcreator/models/mobilenet. Scripts or projects that reference the old CMSIS-NN models or the old MobileNet path will no longer find them; migrate to the TensorFlow Lite workflow or update the model path accordingly.