JeVois  1.18
JeVois Smart Embedded Machine Vision Toolkit
Share this page:
Changes and new features in JeVois 1.8

JeVois 1.8 brings the following new features:

  • Support for the mobile-optimized TensorFlow Lite, which allows one to run TensorFlow deep neural networks on JeVois. See TensorFlowSingle, TensorFlowEasy, and TensorFlowSaliency for example modules using mobilenet and compressed mobilenet pre-trained networks that can recognize 1000 different object categories. These networks can run at up to 83 inferences/s (mobilenet_v1_0.25_128_quant) on the small processor inside JeVois.
  • Bugfix: absolute exposure settings (using setcam autoexp 1 and then setcam absexp 123) can now properly be read back from the sensor (using getcam absexp). Note that the value read back may be truncated depending on frame rate which may limit exposure time.
  • Added a boot option (disabled by default) to connect stream video from 3 JeVois cameras connected to a single USB bus. See JeVois-A33 streaming video from multiple JeVois cameras connected to one USB bus for details.
  • Added an option (enabled by default) to silently drop serial messages when the serial buffer is full. This avoids issues with JeVois appearing to hang when it is actually trying to send serial messages over the USB connection but there is nobody listening to them on the host computer.
  • Changed jevois-cmd host script to use python serial port handling instead of simple shell commands. Should improve reliability in some newer host Linux versions.
  • Added support for operation with no video streaming over USB in ObjectDetect.
  • New boot option: Create an empty file BOOT:quietcmd to suppress sending 'OK' back to the user after every successful command received at the command line. Recommended for advanced users only.
  • New boot option: Create an empty file BOOT:nopython to disable support for modules written in python. When activated, attempting to load a python module will throw an exception. Disabling python saves a lot of memory and may be useful when using C++ modules that run large deep neural networks.