Dec 5, 2023: JeVois 1.20.0 released. On
JeVois-Pro, compile your own C++ modules on device with
no need to install any SDK or to even have a separate desktop computer! It is slow but works great. Updated
to OpenCV 4.8.0, and more improvements. See more
details.
Jun 26, 2023: JeVois 1.19.0 released. Update to OpenCV 4.7.0, OpenVino 2022.3 with TIM-VX
and ARM-Compute backends. Enhanced support for python neural nets with custom post-processing
backends. New support for ONNX Runtime, and more
improvements. See more details. Also
see JeVois-Pro deep learning benchmarks.
Sep 19, 2022: JeVois 1.18.0 released. Update to OpenCV 4.6.0 + OpenVino 2022.1 with TIM-VX
and ARM-Compute backends. New support for 26-TOPS Hailo-8 neural
accelerator, deep learning conversion
tutorials, and more improvements. See more
details. Also see JeVois-Pro deep learning
benchmarks and JeVois + Hailo-8 demo video
May 18, 2022: JeVois 1.17.0 released. Update to OpenCV 4.5.4, Tensorflow 2.5, better handling of incorrect deep nets in DNN module, and more improvements. See more details.
June 30, 2021: Introducing JeVois-Pro. New deep learning smart camera packs a 6-core
processor and up to 13 trillion operations per seconds of deep neural network processing.
See new
product page.
Sept 10, 2020: JeVois 1.15.0 released. Update to OpenCV 4.4.0, Tensorflow 1.15.3 which is
about 25% faster on MobileNets, latest Darknet/NNPACK, fixed GPU/OpenGL-ES support, and more
improvements. See more details.
Sept 13, 2019: JeVois 1.14.0 released. Minor update to OpenCV 4.1.1, new auto-detection of
supported camera sensors, new deep networks, and a few more
improvements. See more details.
Jun 14, 2019: JeVois 1.13.0 released. Major update to OpenCV 4.1.0, buildroot 2019.05,
gcc/g++/gfortran 8.3, boost-1.70.0, eigen-3.3.7, python-3.7.3, python-numpy-1.16.3, new python-scipy-1.3.0
(we got it to cross-compile!), many more pre-installed python packages, and python-pip-19.1 which can be
invoked to install new python packages (for now, only pure-python packages with no compiled dependencies can
be easily installed. We are working on a variant of piwheels to allow one to install any available python
packages, coming soon!) See more details.
Mar 27, 2019: We are back online! Thank you for your patience while our network
infrastructure was being upgraded. We now have a full gigabit connection on jevois.org, which should make
software and data downloads very fast.
Feb 28, 2019: JeVois 1.12.0 released. Upgraded to OpenCV 4.0.1, support for optional 2MP
ov2640 image sensor, support for optional 1.2MP global shutter AR0135 sensor which we coupled with an
ICM-20948 9-axis inertial unit (IMU), new modules for facial emotion recognition, YOLO with int8 inference
(but it is slow...), and more. See more details.
Jan 10, 2019: JeVois Inventor 0.5.0 released. Upgraded to Qt 5.12, support for
dark mode in MacOS Mojave, remember local save directory, new Preferences menu to persistently
select a startup vision module or to start headless.
Dec 7, 2018: JeVois 1.11.0 and JeVois Inventor 0.4.0 released. Upgraded to OpenCV 4.0.0
including all contrib and non-free packages, and deep neural networks (DNN) module. Improved support for
optional AR0135 Global Shutter sensor (color and monochrome) and for ICM-20948 9-axis inertial measurement
unit. Bugfixes and more. See more details.
Oct 25, 2018: JeVois 1.10.0 and JeVois Inventor 0.3.0 released. Upgraded to OpenCV
4.0.0-beta including all contrib and non-free packages, and deep neural networks (DNN) module. New JeVois
module
PyClassificationDNN (programmed in
Python!) runs SqueezzeNet to recognize 1000 object types, and can load your own deep nets created with
Caffe, TensorFlow, Darknet, Torch or ONNX. See more
details.
Aug 3, 2018: JeVois 1.9.0 released. Upgraded to OpenCV 3.4.2 including all contrib packages
and deep neural networks (DNN) module. New JeVois modules
DetectionDNN and
PyDetectionDNN (programmed in Python!)
run Darknet-YOLO v3, MobileNet v2 + SSD, OpenCV Face Detection network, and more deep nets created with
Caffe, TensorFlow, Darknet or Torch. See more details.
Jul 19, 2018: JeVois Inventor beta 0.2.0 released. Added line numbers to all editors,
spinboxes to show/edit values of all slider controls, enhanced camera detection on Windows platforms, links
to online docs and help. See JeVois Inventor doc for details.
Jun 21, 2018: JeVois 1.8.1 and JeVois Inventor beta 0.1.0 released. JeVois Inventor is a
new graphical front-end for JeVois which makes programming new machine vision pipelines using full
Python 3.6 + OpenCV 3.4 easier than ever before. See JeVois Inventor
doc for details.
Apr 6, 2018: New lens options available for JeVois: NoIR, 90deg no-distortion, and 120deg
fisheye. Very limited quantities in stock now. See JeVois doc
and JeVois Inc website for details.
Mar 30, 2018: JeVois 1.8.0 released with TensorFlow support and more. Includes
mobile-optimized TensorFlow Lite, which allows one to run TensorFlow deep neural networks on JeVois. See
TensorFlowSingle,
TensorFlowSaliency, and
TensorFlowEasy for example modules
using compressed mobilenet pre-trained networks that can recognize 1000 different object categories at
up to 83 inferences/s on JeVois. See more details.
Feb 21, 2018: JeVois 1.7.2 released
with support for up to 3 cameras on a single USB bus
(using reduced bandwidth for each camera), default drop policy on serial-over-USB port (so that serial
messages will not block processing if nobody is listening), and fixed absolute exposure settings.
Jan 24, 2018: JeVois 1.7.1 released with an example of parallel processing in
Python. Have a look at the experimental new
module PythonParallel if you are
interested in Python multi-processing on JeVois.
Jan 18, 2018: JeVois 1.7.0 released with OpenCV 3.4 and more. Includes new microSD image
and new Debian packages for Ubuntu 17.10 and 16.04 (Ubuntu 17.04 is no longer
supported). See more details.
Jan 11, 2018: JeVois 1.6.3 released today includes a new
Python keypoint-based object detection
module as an example of machine vision for the First Robotics Competition (FRC) 2018 Power Up (sm)
challenge, including 6D object pose estimation in Python.
Jan 9, 2018: JeVois 1.6.2 released today includes a new
Python shape-based object detection
module which shows how to do full 6D object pose estimation in Python.
Jan 3, 2018: Happy new year everyone! JeVois 1.6.1 released today includes a new
FIRST Robotics example vision module
which may interest FRC teams looking for a programmable vision solution for their robot. See the
companion tutorial to get started.
Dec 11, 2017: Fresh from the NIPS workshop on
metalearning, born again neural networks are how one can improve
over time the performance of tiny deep neural networks like those we run on JeVois!
Dec 8, 2017: JeVois 1.6.0 released. Fixed issues with SDK Debian packages for Ubuntu
16.04, added some workarounds to handle bogus Windows 10 requests to the camera, added support for
Ubuntu 17.10, updated to the latest darknet-nnpack, and more,
see more details.
Oct 24, 2017: New JeVois guided tour. An interactive tutorial that takes you through
many of the available machine vision modules of JeVois in a simple and intuitive way. Ideal to try with
kids!
Get it from JeVois Start
Oct 19, 2017: JeVois and Prof. Itti selected by ARM as one of 10 worldwide for new ARM
Innovators program. ARM Innovators are key technical experts helping to extend ARM architecture and
to share their knowledge into a wide variety of exciting areas. See the announcement
on hackster.io
Oct 18, 2017: JeVois 1.5.2 released with Saliency+Darknet, YOLO resizing.
- Saliency + Darknet: Detect the most interesting (salient) object in the scene, then recognize it
using a deep neural network. Saliency detection runs at 15 frames/s at 320x240 or 10 frames/s at
640x480 (limited to not consume too much CPU, saliency can actually run alone at 60+ frames/s at
320x240), and neural network recognition (1000 possible object categories) runs at about
200ms/object (about 5 objects/s).
See DarknetSaliency Module.
- Darknet and YOLO runtime resizing: the neural networks used in
modules DarknetSaliency,
DarknetSingle
and DarknetYOLO can now be resized
at runtime. This allows you to trade off between speed and accuracy. Smaller network sizes run
faster but are less accurate and have more difficulty identfiying small objects.
Sept 29, 2017: JeVois 1.5 released with OpenCV 3.3, g++ 7.2 and more.
- JeVois is now based on buildroot 2017.08 and includes OpenCV 3.3.0, python 3.6, boost 1.64, eigen
3.3.3, latest tiny-dnn and darknet-nnpack deep neural networks, and many other latest package
versions.
- This update changes the JeVois operating system. Developers need to update to the new debian
package jevois-sdk-dev-1.5.0 (and its dependencies) released for Ubuntu 17.04 and 16.04
at jevois.usc.edu. Also see
Setting up for programming
JeVois.
- And more, see more details.
Sept 7, 2017: JeVois 1.4 released with new deep neural networks. Object recognition on
ImageNet (1000 object categories) and Pascal VOC using Darknet and Darknet YOLO. Not fully optimized
for speed yet but already looking very promising.
See two new modules in jevoisbase:
- DarknetSingle: passes a central
224x224 crop to a Darknet neural network. The neural network precicts the top N object category
candidates (among 1000 possible from imagenet), and those above a minimum confidence threshold are
shown and sent to serial.
- DarknetYOLO: passes a whole
640x480 frame to a Darknet YOLO (you only look once) network that combines detection (finding
where objects might be) and identification (what objects are) for all objects in the scene in
one pass.
- And more, see more details.
July 13, 2017: Setup Facebook and Twitter. See the icons at the top of this page. We
will from now on post links to new tutorials, new documentation pages, and new software versions on
these channels.
July 12, 2017: JeVois 1.3 released.
- Standardized serial messages for greater inter-operability between JeVois and Arduino.
- New host command jevois-add-videomapping to add a new mapping to videomappings.cfg,
checking for duplicates.
- New host command jevois-cmd to send a command to JeVois and get the result without having
to open a serial terminal and connect it to JeVois.
- New host command jevois-usbsd to instruct a connected JeVois camera to start/stop
exporting its microSD as a virtual flash drive without having to open a serial terminal app.
- New docs, including on using in low light.
- New tutorials for C++ and Python
- And more, see more details.
Get latest
microSD image
July 1, 2017: JeVois 1.2 released.
- Compile and install new modules directly to live microSD inside JeVois
- More documentation and tutorials
- Full stack trace is now reported with Python errors
- New video error reporting, showing a exception messages in the video stream to ease debugging of
modules in c++ and Python
- Added support for JeVois parameters of types std::pair, cv::Size and cv::Point
Get latest
microSD image
June 16, 2017: JeVois 1.1 released.
- Access the microSD card inside JeVois live over USB as a virtual USB flash drive. Directly
edit config files, add training data, retrieve data recorded by JeVois, or install/update machine
vision modules without having to physically take out the card.
- Program machine vision modules in Python 3.5 with full Python-OpenCV 3.2.0 support. Combined
with the live access to microSD over USB, edit a computer vision module written in Python directly
on the JeVois camera without having to remove the card.
- New suite of Ubuntu 17.04 deb packages for developers, should (hopefully!) eliminate headaches
with trying to install the jevois-sdk.
- Access a Linux login shell on JeVois, either over the hardware 4-pin serial port, or using
serial-over-USB, while machine vision software is running.
- New camera settings for low-light and night vision, using automatic reduction of framerate to
increase exposure time.
- And more, see more details.
Get latest
microSD image
