JeVois 1.18 brings the following new features:

New support for Hailo-8 neural accelerator, up to 26-TOPS in a small M.2 A+E 2230 board that can fit inside JeVois-Pro! This is by far the fastest neural accelerator available to date, and it can run larger models like YOLOv5m or YOLOv7 are very nice framerates. See SPU entries in JeVois-Pro Deep Neural Network Benchmarks
Updated to OpenCV 4.6.0, OpenVino 2022.1
Support for TIM-VX library to execute models on the JeVois-Pro integrated NPU directly via OpenCV (no conversion to NPU blob format needed, load model directly from ONNX, etc; model still should be quantized to int8). See TIM-VX and OpenCV+TIM-VX.
Support for Intel Myriad-X emulation on ARM CPU through the ARM Compute Backend to OpenCV. See ARM CPU Plugin and github repo. Load any OpenVino model for VPU and try them out using CPU-based emulation even if you do not have a VPU hardware dongle. Emulation is quite slow compared to the actual VPU chip, though. See VPU vs VPUX in JeVois-Pro Deep Neural Network Benchmarks
Core JeVois: new support for dynamic JeVois parameters. These are strongly typed parameters that a can be accessed via command-line, console, or GUI.
Python: Allow creation of and access to JeVois parameters in Python modules. Have a look at PythonTest for an example. Python modules can now have tunable parameters that users can adjust in the GUI, just like C++ modules.
Python: Added support for Python Components. These are JeVois Component objects that can be assembled into component hierarchies just like C++ components, but they are written in python. Python components can hold JeVois parameters created from Python.
Core JeVois: moving towards using std::filesystem for file and paths operations
DNN: support for pre-processor, network, and post-processor written in python. This will make it much easier to add new network types, as often pre-processing and post-processing code in Python is provided by the network designer. These can now be easily ported to JeVois. Check out:
- PyPreBlob.py - sample pre-processor in Python
- PyNetOpenCV.py - load network and run inference using the OpenCV API
- PyPostClassify.py - report top-scoring classes
DNN: better error messages when expected network inputs/outputs differ from what is received.
DNN: better and more general pre-processing. Ability to show lots of details when PreProcessor::details is checked in the GUI.
DNN: much improved RAW YOLO preprocessing, about 4x faster, supports various yolo shapes and both old/new coordinate styles. Better handling of YOLO anchors.
JeVois-Pro GUI: Improved text editor, new jevois::GUIeditor class. Added file browser and support for keyboard shortcuts. Can now edit any file on disk and create new files if needed.
OpenGL GUI: added d2i() to convert coordinates from screen to displayed image. See example of use in new GUI version of ColorFiltering
JeVois-Pro: support for 4K display. Can be enabled by editing /boot/env.txt, not enabled by default as it slows down the overall operation a bit, likely because of higher traffic on the memory bus.
JeVois-Pro: More deep learning models in models.yml model zoo. Split off model.yml into separate child files for npu, tpu, vpu, etc. Added support for global paremeters that can be set and unset. New zoo files are much easier to read and edit now. Zoo files are directly accessible undr the Config tab of the JeVois-Pro GUI.
JeVois-Pro: Better fan control. Note: as there currenty is no Hailo temp available in /sys/, we always turn on the fan to 100% speed if a Hailo-8 accelerator is installed in JeVois-Pro.
New docs and tutorials on how to convert deep networks for the various accelerators available on JeVois-Pro:
- Running neural networks on JeVois-A33 and JeVois-Pro
- Converting and Quantizing Deep Neural Networks for JeVois-Pro
- Running neural networks on JeVois-A33 or JeVois-Pro using OpenCV
- Converting and running neural networks for JeVois-Pro NPU
- Converting and running neural networks for Hailo-8 SPU
- Converting and running neural networks for Coral TPU
- Converting and running neural networks for Myriad-X VPU
replaced frameNum() in JeVois Module by a more widely accessible jevois::frameNum() free function. Reports the current video frame number. Reset to 0 when a new machine vision module is loaded.
added new usilities such as jevois::joinall(), jevois::num2str(), jevois::secs2str(), Hailo-related helpers, quantization-related, etc
New LTHROW() macro is like LFATAL() but does not show the error in the console, just throw it. Useful in sub-routines where the caller will catch any exception and possibly reformat it to provide unified error messages. Just like a plain throw, except that the same JeVois streaming syntax as in LDEBUG(), LINFO(), etc can be used.
Miscellaneous bug fixes and performance improvements.