JeVois  1.20
JeVois Smart Embedded Machine Vision Toolkit
Share this page:
Changes and new features in JeVois 1.19

JeVois 1.19 brings the following new features:

  • New support for ONNX Runtime framework on CPU, both in C++ and in Python. This allows you to run ONNX networks on CPU, with greater compatibility than the OpenCV framework provides. See new ORT pipeline type in the DNN module. Added several sample networks.

  • Improved ability to write pre-processor, network, and post-processor in Python for the DNN module. Provided several new examples in jevoisbase:

    • share-jevoispro/pydnn/net/PyNetOpenCV.py - load a network and run inference using OpenCV
    • share-jevoispro/pydnn/net/PyNetORT.py - load a network and run inference using ONNX Runtime
    • share-jevoispro/pydnn/net/PyNetURetinex.py - ariant of PyNetORT.py with extra fixed input for this network
    • share-jevoispro/pydnn/post/PyPostClassify.py - simple object classification post-processor in Python
    • share-jevoispro/pydnn/post/PyPostDAMOyolo.py - object detection post-processor for DAMO-YOLO networks
    • share-jevoispro/pydnn/post/PyPostDepth.py - display depth map as semi-transparent overlay
    • share-jevoispro/pydnn/post/PyPostStub.py - does nothing, useful as a starting point for new network types
    • share-jevoispro/pydnn/post/PyPostURetinex.py - display partial overlay with user mouse interaction example
    • share-jevoispro/pydnn/post/PyPostYolo.py - wrapper around our very fast C++ YOLO post-processor

    Remember that because python is not re-entrant or parallelizable, if more than one of pre-processing, network inference, or post-processing are in python, that will force synchronous DNN processing, which may lower the framerate of the GUI on slow networks.

  • Updated to OpenCV 4.7.0, OpenVino 2022.3, EdgeTPU 1.16, PyCoral 2.0.0, HailoRT 4.13.0
  • Core JeVois: Fixed selection of default videomapping from file, updated contribs to include onnxruntime, cleaned up some include and dependency paths in CMakeLists.txt.
  • Core JeVois: Added more Python bindings for getNumInstalledTPUs(), getFanSpeed(), etc
  • Movie input: added support for dual-stream in JeVois-Pro (where the input at one resolution is sent to display, while the same image at another resolution is sent to processing). To use pre-recorded movie inputs instead of live video, use jevoispro.sh --cameradev=/path/to/movie.mp4. Also fixed movie rewinding and looping.
  • Programmers: created new github repo mydnn as an example standalone clone of the JeVois DNN module. You can modify it, cross-compile it, and install it on your camera.
  • DNN: PreProcessorBlob is now only returning 1 blob by default, which can be changed by parameter numin (number of inputs to send to the network). Then, new Network parameter extraintensors can be used to specify extra fixed inputs. This is used in the URetinexNet example, as this network expects both in image input, and a single float input that sets an exposure parameter.
  • DNN: Added i2b() in PreProcessor to allow converting coordinates from input image to blob, which is useful when converting mouse coordinates first from display to image (using GuiHelper::d2i()), and then from image to blob (using new PreProcessor::i2b()). Both d2i() and i2b() correctly handle letterboxing, scaling, streching, etc. See example of use in PyPostURetinex.py
  • DNN: Improved semantic segmentation post-processor to allow for a threshold (only values above threshold will be displayed as semi-transparent OpenGL overlay).
  • JeVois-Pro: improved crispness of the ImGui rendering, which now looks super nice even at GUI scale 1 (under the Parameters tab of the GUI, enable "Show Systems Parameters", and then in the section on Graphical User Interface, play with the gui scale parameter).
  • JeVois-Pro: Added more python wrappers for ImGUI, to allow users to check for mouse coordinates, button clicked, etc in python. Example of use in PyPostURetinex.py
  • JeVois-Pro: support for sending serial outputs to USB using gadget serial. The gadget can be enabled in the JeVois-Pro GUI under the System tab, after which module outputs and/or logs can be sent to the mini-USB port operating in device mode. You can then connect JeVois-Pro to a host computer using a regular USB-A to mini-USB cable, and the JeVois-Pro serial console will be accessible on host via /dev/ttyACM0.
  • Serial-over-USB: Now can handle USB cable disconnects/reconnects, by trying to re-open the device after the connection was lost. Also better handling of USB send buffer overflows while the cable is disconnected.
  • JeVois-Pro: More deep learning models, including several new ONNX-Runtime models. Added fields for comment and URL to the model zoo entries.
  • JeVois-Pro: Better fan control, including temperature monitoring on Hailo-8 accelerator board.
  • JeVois-Pro: Better module info display with wrapping of bullet text.
  • JeVois-Pro: New benchmark parameter in Pipeline to run an automated benchmark that cycles through all available networks according to the current pipeline filter. Benchmark stats are saved in HTML table format to /jevoispro/share/benchmarks.html
  • JeVois-Pro: We have working prototypes of 8MP/4K OS08a10 sensor boards, and 2.3MP global shutter AR0234CS. The color calibration is not good on the AR0234, will need to tweak the ISP calibrations in the kernel. Please let us know if you are interested in these sensors. Added camera sensor auto-detection based on which kernel overlay is active.
  • Miscellaneous bug fixes and performance improvements.