Dev Kit Weekly: Maxim Integrated MAXREFDES178 Cube Camera Reference Design
October 21, 2022
At least for today, that is simply untrue, because this week’s device is capable of not only face identification, but also face classification and keyword spotting. And it’s just an added bonus that it’s the cutest little cube you ever did see. This tiny AI machine is Maxim Integrated’s MAXREFDES178 cube camera reference design.
The MAXREFDES178 may look unassuming, but it packs quite a punch, starting with the dual MAX78000 integrated circuit microcontroller boards nestled inside. These boards are designed specifically for AI-powered image and audio applications, hence the double usage. To allow the MAXREFDES178 to run both image- and audio-based AI models at the same time, the microcontrollers include Arm Cortex-M4 processor with a floating point unit as well as CNN acceleration hardware that concentrates on image and video data from an onboard image sensor plus voice and audio from the connected digital microphone via a low-power audio codec. But more on the CNN accelerator later.
There’s also a 32-bit RISC-V coprocessor that can run at up to 60MHz.
Each board also includes 512KB of internal QSPI flash memory, as well as an additional 128KB of internal QSPI SRAM. The MCUs are optimized for running deep convolutional neural networks with a hardware-based accelerator designed for battery powered devices like the MAXREFDES178 to ensure that power consumption remains minimal (we’re talking microjoule-level) executing inferences.
While all the features and functionality on the MAX78000 MCUs are great, we haven’t even gotten to the connectivity yet. No, we have an entire third board dedicated to that. The connectivity board is based on the dual core MAX32666, which also leverages the 32-bit Arm Cortex-M4F, but that makes up both processing cores in this MCU. One of these cores, Core0, is responsible for handling the MAXREFDES178’s USB and Bluetooth Low Energy software stacks. The other, Core1, controls the device’s main application logic and the transferring of audio, image, and control data between other components like the audio codec, memory, on-chip BLE radio, and so on.
The MAXREFDES178’s connectivity board provides both wired and wireless communication between device and user through the littlest capacitive LCD touch screen you ever did see, and a ceramic antenna connected to the connectivity board’s MAX32666 enables communication to external devices like tablets, phones, or computers. The connectivity board also houses a MAX20203 power management IC, connected via I2C bus, that charges the camera and battery when connected via USB, or uses buck and boost converters to generate the required voltage.
Not sold yet? The MAXREFDES178 also comes with its very own MAXDAP-TYPE-C debug adapter, which includes multiple USB ports to connect to the device itself, enable debugging and programming, and assist in data transfer.
All that hardware is awesome, obviously, but what does it look like in action? I spoke about face identification at the beginning of the video, and it just so happens that the MAXREFDES178 ships pre-loaded with a low-power convolutional neural network inference engine that can run AI computations on the IoT edge. And to turn this thing into more than just a paperweight, the kit comes packed with a Face Identification demo from Maxim. The demo is based on the FaceNet multitask-cascaded convolutional neural network model that can be optimized and trained in either PyTorch or TensorFlow-Keras environments using data sets like VGGGFace-2 and YouTubeFaces. The CNN performs a series of actions on subject images including facial extraction, alignment, and, of course, identification.
The FaceNet model is extremely popular in facial identification applications, and the demo tests the accuracy and performance of the model against the MaximCeleb dataset, a group of 15 male and 15 female celebrities. However, the model contains 7.5 million parameters and requires 1.6G floating point operations, which, as you may have guessed by the limited amount of memory on-chip and onboard, isn’t really optimized for IoT devices like the MAXREFDES178 kit.
Fortunately, the model in the demo is just based on FaceNet, but isn’t FaceNet. It leverages a process called “knowledge distillation” that helps transferring large models to smaller models and is tuned for the MAX78000 microcontroller, resulting in the AI85FaceIdNet model. A synthesizer on the MAX78000 automatically generates C code and an API that allows for easy integration of inference outputs into your applications.
More on that can be found in an application note from Maxim Integrated.
The MAXREFDES178 kit can be yours for just $169 from the Maxim online store. That includes the cube with LCD, enclosure, and battery, the debug adapter, and of course, access to the aforementioned demo. But there is another way. You can enter the raffle below for a chance to win this kit for free. We’ll even pay the shipping.
Thanks for tuning in, good luck in the raffle, and we’re looking forward to seeing your faces on the next Dev Kit Weekly. Remember, we’ve got this MAXREFDES178, so something might be watching.