Powerful JeVois-Pro Packs a Full Ubuntu System Into an AI-Ready Camera Module Weighing Just 2.8oz
Powerful yet compact, this camera module is a fully standalone device — connect a keyboard, mouse, monitor, and start developing.
Educational machine vision startup JeVois has launched a crowdfunding campaign for a compact deep learning camera — which, it promises, is fully open source and programmable in C++ or Python.
"Unlike a traditional camera that relies on a human viewer to interpret the contents of what it sees, JeVois-Pro runs powerful machine vision and deep learning algorithms right inside the camera," says company founder Laurent Itti, PhD, professor of computer science, psychology, and neuroscience at the University of Southern California.
"This allows you to directly connect JeVois-Pro to a custom robot, home automation system, drone, surveillance system, etc. You let JeVois-Pro handle the machine vision while you focus on developing the high-level functionality of your system."
The compact camera system is built around an Amlogic A311D system-on-chip with four Arm Cortex-A73 cores running at 2.2GHz, two Cortex-A53 cores running at 1.8GHz, and a quad-core Mali-G52 graphics processor running at 800MHz. On top of this is a neural coprocessor offering a claimed five trillion operations per second (TOPS) of compute, 4GB of LPDDR4 RAM, and a TDK InvenSense ICM-20948 inertial measurement unit with three-axis accelerometer, gyroscope, and compass.
The camera side of the system is handled by a user-swappable Sony IMX290 Starvis sensor module with backside illumination, a 1/2.8" 12mm lens, wide dynamic range support, and up to 1920x1080 capture at 120 frames per second. For output, there's an HDMI 2.1 port, a microSD slot for storage, two USB Type-A ports, a mini-USB On-The-Go (OTG) port, and a custom sensor connector for adding up to two additional camera modules.
If the on-board processing potential lacks power for your project, JeVois-Pro's design also features an M.2 E-key slot for PCI Express add-ons —including, the company promises, compatibility with Google's Coral TPU accelerator. The slot can also be used for a Wi-Fi or other network card, or the company's custom extension for eMMC flash storage.
Running Ubuntu 20.04 LTS and coming pre-loaded with OpenCV, OpenVino, and the JeVois Core Library — boasting, the company says, more than 30 machine vision modules — the camera supports a range of frameworks including TensorFlow Lite, Caffe, ONNX, MxNet, and Darknet.
A major selling point of the device is its open nature: "The operating system, firmware, and JeVois machine vision framework that run on the device are fully open source," Itti promises, "except for a few low-level hardware drivers (e.g., MALI GPU drivers - but open source variants often exist even for those). Source code for Linux kernel, U-Boot bootloader, JeVois framework, and JeVois machine vision algorithms is freely available on GitHub. Full [hardware] schematics will be released at [the] end of [the] campaign."
Despite being, in effect, a surprisingly powerful Arm-based computer with desktop-class capabilities, the JeVois-Pro is surprisingly lightweight at 2.8oz with the bundled case, fan, and heatsink — making it suitable, its creator claims, for weight-sensitive projects like drone installation.
The JeVois-Pro is now funding on Kickstarter, with rewards starting at $379 for a starter kit including the camera, a free Coral Edge TPU accelerator adding an extra 4 TOPS of compute, a 64GB microSD card, micro-HDMI to HDMI adapter, 12V power supply, and 15cm UART, general-purpose input/output (GPIO), and auxiliary power cables. Delivery is scheduled for September this year — an aggressive timescale.