Earlier this year Google introduced the AIY voice project, which allowed makers to turn a Raspberry Pi into a voice-controlled assistant, using the Google Assistant SDK.
The company has now launched the AIY (AI yourself) Vision Kit that lets you turn Raspberry Pi equipment into an image-recognition device. The kit is powered by Google's TensorFlow machine-learning models and will soon gain an accompanying Android app for controlling the device.
According to Google, Vision Kit features “on-device neural network acceleration”, allowing a Raspberry Pi-based box to do computer vision without processing in the cloud. The AIY Voice Kit relies on the cloud for natural-language processing.
As with the AIY Voice Kit, makers will need to provide their own Raspberry Pi Zero W, a Raspberry Pi camera, a 4GB SD card, and power supply. The Vision Kit itself includes the VisionBonnet accessory board, cables, the cardboard box and frame, lens equipment, and a privacy LED to tell others when the camera is on.
The VisionBonnet board was developed by Google and is powered by Intel's Movidius MA2450 vision processing chip. This chip is the secret sauce of the Vision Kit.
Kai Yick, a member of the AIY Projects team, told the official Raspberry Pi magazine MagPi that the chip is 60 times faster at performing computer vision than if it were to rely on a Raspberry Pi 3. The VisionBonnet is connected to the Raspberry Pi Zero W by a cable supplied in the kit.
Vision Kit makers can use several neural-network programs, the first being one that can detect when people, cats, and dogs are in view, according to AIY Projects. Another neural network will detect happiness, sadness, and other sentiments. A further program, based on MobileNets, can detect 1,000 different objects such as a chair, orange, or cup.