Vision for the Future

A round-up of some Edge Impulse-friendly computer vision boards to help you cut through the noise and pick the best option for your needs.

Nick Bild
21 days agoMachine Learning & AI

Whether you are designing an algorithm that helps a self-driving car plan out a safe path of travel, or you want to create your own gesture recognition system to control your smart home, you will need to get acquainted with the best available computer vision boards to suit your unique requirements. With the help of the Edge Impulse team, we searched far and wide to round up a selection of boards that are big, small, and everywhere in between. Some of the hardware is minimalistic, and some comes with all the bells and whistles.

We put each and every one of these boards through its paces so that you don’t have to. Read on to learn the pros and cons of each, and to find the perfect choice for your application.

Espressif ESP-EYE

Image recognition for AIoT applications is this tiny board’s specialty. If you want to go very far outside of this area, you might need to look at other options. Integration with other hardware is challenging as the ESP-EYE does not expose any GPIO pins. For the price, it is exceptionally well made. In our testing, the Wi-Fi radio is more reliable, and has a better range, than the ESP32-CAM.

  • Cost: ~$20.00
  • Processing: ESP32 microcontroller
  • Memory: 8 MB PSRAM
  • Storage: 4 MB Flash
  • Cameras: 1 included, 2-megapixel
  • Connectivity: Wi-Fi, Bluetooth
  • Power Consumption: ~1 watt
  • Size: 21 mm x 41 mm
  • Bonus Feature: A microphone plus programmable buttons and LEDs
  • Edge Impulse: Fully supported in the Edge Impulse Studio,so you can deploy your projects directly to the board with native firmware downloads.

ESP32-CAM

The ESP32-CAM is useful for image recognition tasks, much like the ESP-EYE, although it has less memory, is somewhat larger, and generally lacks the polish of the ESP-EYE. But considering the ultra-low price, and the flexibility offered through the exposed GPIO pins, it is worth having at least one of these in your parts bin at all times.

  • Cost: ~$8.00
  • Processing: ESP32 microcontroller
  • Memory: 520 KB SRAM, 4 MB PSRAM
  • Storage: 4 MB Flash
  • Cameras: 1 included, 2-megapixel
  • Connectivity: Wi-Fi, Bluetooth
  • Power Consumption: ~1 watt
  • Size: 27 mm x 40.5 mm
  • Bonus Feature: SD card reader
  • Edge Impulse: Not an “official” Edge Impulse board, but it definitely works! You can check out this blog post to get started using the ESP32-CAM to run your computer vision models.

NVIDIA Jetson AGX Xavier

If you have heavy AI workloads to deal with, the Jetson AGX Xavier should definitely be on your short list. With 32 teraflops of computing power and just about every feature you can conceive of, this module will make short work of whatever you throw at it.

  • Cost: $999.00
  • Processing: 8-Core Carmel Arm v8.2 64-Bit CPU / NVIDIA Volta architecture GPU with 512 NVIDIA CUDA cores and 64 Tensor cores
  • Memory: 32 GB
  • Storage: 32 GB eMMC 5.1
  • Cameras: Supports up to 6 CSI cameras
  • Connectivity: 10/100/1000 BASE-T Ethernet
  • Power Consumption: 10-30 watts
  • Size: 100 mm x 87 mm
  • Bonus Feature: 2 x 7-Way VLIW Vision Processor
  • Edge Impulse: The AGX Xavier is quite similar, but much more powerful, than the Jetson Nano, which is an officially supported device. So, by following the Jetson Nano instructions, you should be able to deploy the Linux runner on it just fine.

Useful Sensors Person Sensor

This is a pre-programmed unit designed to do one thing: detect people. If that is not what you are looking for, or if you want to experiment with your own machine learning algorithms, this is not the board you are looking for. But if you want to detect people without worrying about the details, there is no easier way to do it.

  • Cost: $9.95
  • Processing: Not specified
  • Memory: Not specified
  • Storage: N/A
  • Cameras: 1 included
  • Connectivity: N/A
  • Power Consumption: 0.5 watts
  • Size: 20 mm x 20 mm
  • Bonus Feature: Qwiic interface for easy integration with other devices
  • Edge Impulse: The Useful Sensors is brand new at the time of this writing, so there is no Edge Impulse support just yet. But be sure to check out this livestream with the Useful Sensors team to learn more about their plans.

Texas Instruments TDA4VM

Robots, smart cameras, and intelligent machines of all sorts are within reach with this powerful platform. It provides 8 teraflops of compute power, and yet with TI’s extensive documentation, you can have a vision-based prototype up and running in under an hour. The TDA4VM evaluation kit can run Linux, so standard frameworks like TensorFlow Lite can be leveraged in your designs.

  • Cost: $249.00 (starter kit)
  • Processing: 2 x Arm Cortex-A72 CPUs operating at up to 2 GHz, 6 x Arm Cortex-R5F MCUs operating at up to 1 GHz
  • Memory: 4 GB
  • Storage: eMMC 5.1 and SD3.0 interfaces
  • Cameras: Up to 8
  • Connectivity: Ethernet
  • Power Consumption: ~15 watts
  • Size: 81.44 mm x 44.77 mm
  • Bonus Feature: Deep learning and vision processing accelerators
  • Edge Impulse: Official support for the TI TDA4VM is planned and currently being worked on. In the meantime, the standard Linux support will work, as outlined here.

Arduino Nicla Vision

This super small board is a great choice for industrial applications including asset tracking, image detection, object recognition, and predictive maintenance. It comes loaded with sensors to power your creations when you need more than just vision. With support from the Arduino community and easy-to-use tools like Arduino IDE, developing prototypes is frustration free.

  • Cost: $115.00
  • Processing: 1 x Arm Cortex-M7 at up to 480 MHz, 1 x Arm Cortex-M4 at up to 240 MHz
  • Memory: 1 MB RAM
  • Storage: 16 MB QSPI Flash
  • Cameras: 1 onboard, 2 megapixel
  • Connectivity: Wi-Fi, Bluetooth Low Energy
  • Power Consumption: Not specified
  • Size: 22.86 mm x 22.86 mm
  • Bonus Feature: 6-axis IMU and microphone
  • Edge Impulse: The Nicla Vision is a fully supported Edge Impulse device, you can download ready-made firmware direct from the Studio.

Arduino Portenta H7 + Portenta Vision Shield

The Portenta H7 is targeted towards all manner of smart projects, from home automation to smart gardens. It is compatible with the Arduino IoT Cloud, which makes it easy to visualize data and control your projects from anywhere in the world. This platform is also suited for industrial and laboratory applications.

  • Cost: $113.90 (H7), $51.80 (Vision Shield)
  • Processing: STM32H747XI dual Cortex-M7+M4 32-bit low-power Arm MCU
  • Memory: 8 MB SDRAM
  • Storage: 16 MB NOR Flash
  • Cameras: 1 onboard, 320 x 320 pixel resolution
  • Connectivity: Ethernet
  • Power Consumption: Not specified
  • Size: 66.04 mm x 25.4 mm
  • Bonus Feature: Chrom-ART graphical hardware accelerator
  • Edge Impulse: Similar to the Nicla Vision, the Portenta H7 and it’s Vision Shield are officially supported Edge Impulse devices, you can get started with this blog post and embedded video.

Google Coral Dev Board Micro

This is a development board designed for low-power systems with fast on-device inferencing for vision applications. It has high-density connectors for easy expansion via add-on boards that make the platform highly adaptable to different use cases. If you prefer to build machine learning models with TensorFlow Lite, then this board might be for you. The Edge TPU was created specifically to accelerate TensorFlow Lite algorithm execution.

  • Cost: $79.99
  • Processing: NXP i.MX RT1176 (Cortex-M7 at 800 MHz, Cortex-M4 at 400 MHz), Coral Edge TPU
  • Memory: 64 MB
  • Storage: 125 MB Flash
  • Cameras: 1 onboard, 324 x 324 pixel resolution
  • Connectivity: N/A
  • Power Consumption: ~3 watts peak
  • Size: 65 mm x 30 mm
  • Bonus Feature: Microphone and GPIOs
  • Edge Impulse: The Coral Dev Board Micro is not actually launched yet, and is only available to pre-order at the time of this writing. However, some documentation is already posted, and it looks like it can make use of the Arduino programming language, so adding Edge Impulse Arduino libraries should be possible.

OpenMV Cam H7

The OpenMV Cam H7 is designed from the ground up with hobbyists in mind. It is programmable with MicroPython, which makes it much easier to deal with the complexities of machine learning algorithms. It also has lots of available GPIO pins to incorporate the device into larger builds. If you have an idea in mind for a computer vision application, but are not quite sure how to get started, you should consider giving this board a try.

  • Cost: $65.00
  • Processing: STM32H743VI Arm Cortex-M7 CCPU at 480 MHz
  • Memory: 1 MB SRAM
  • Storage: 2 MB Flash
  • Cameras: 1 onboard, up to 640 x 480 pixel resolution
  • Connectivity: N/A
  • Power Consumption: ~0.5 watts
  • Size: 35.56 mm x 44.45 mm
  • Bonus Feature: microSD card reader
  • Edge Impulse: The OpenMV Cam H7 is another fully supported platform that is integrated into the Studio, and built firmware can be downloaded and imported directly into the OpenMV IDE.

Raspberry Pi 4

While the Raspberry Pi 4 is not a computer vision board per se, it is just too good to leave out. The Raspberry Pi is a single board computer that packs a punch, so it is quite capable of handling image recognition, object detection, and other vision algorithms with high frame rates. Since it runs Linux, it is extremely versatile in the applications it can be used for, and common machine learning toolkits can be used with it. Given the low price point and utility, this is simply a must have device.

  • Cost: $45.00 - $75.00
  • Processing: Quad-core Arm Cortex-A72 running at 1.5 GHz
  • Memory: 2 GB, 4 GB, or 8 GB
  • Storage: microSD card (not included)
  • Cameras: CSI or USB camera compatible
  • Connectivity: Wi-Fi, Bluetooth, Gigabit Ethernet
  • Power Consumption: ~12.5 watts peak
  • Size: 85 mm x 56 mm
  • Bonus Feature: Dual micro HDMI ports supporting two 4K displays
  • Edge Impulse: One of the most popular boards for developers and makers, the Raspberry Pi 4 (and earlier models as well) are officially supported via the Linux deployment options.

Sony Spresense + Camera Board

When low power consumption is critical, the Spresense is hard to beat. The Spresense’s microcontroller is very efficient, which allows the board to offer great performance while slowly sipping power. Industrial IoT, agriculture tech, and smart city applications are ideal for this platform.

  • Cost: $55.00 (main board), $35.00 (camera board)
  • Processing: 6 x Arm Cortex-M4F CPU at 156 MHz
  • Memory: 1.5 MB SRAM
  • Storage: 8 MB Flash
  • Cameras: 1 onboard, 2608 x 1960 pixel resolution
  • Connectivity: N/A
  • Power Consumption: ~0.1 watts
  • Size: 50 mm x 20.6 mm
  • Bonus Feature: Integrated GPS receiver
  • Edge Impulse: Sony’s Spresense platform is natively supported by Edge Impulse, so deployment is as easy as building and downloading the firmware, then flashing it to the board. Here’s a great deep dive to get you going.

Seeed Studio Grove - Vision AI Module

Seeed Studio has a reputation for bringing cutting edge technologies to the prototyper at a low price point, and this board is no exception. It comes preinstalled with a people detection model, as well as algorithms for several other common use cases. The Grove connector can be used to connect to other modules in the ecosystem and add in numerous additional functionalities without a hassle.

  • Cost: $25.99
  • Processing: Himax HX6537-A at 400 MHz
  • Memory: 1472 KB system memory
  • Storage: 32 MB SPI Ultra Low Power Flash
  • Cameras: 1 onboard, up to 1600 x 1200 pixel resolution
  • Connectivity: N/A
  • Power Consumption: Not specified
  • Size: 40 mm x 20 mm
  • Bonus Feature: Accelerometer and microphone included
  • Edge Impulse: The Seeed Studio Vision AI Module is brand new, just launching a few weeks back. Edge Impulse support is coming soon, but you can get a sneak preview and see a project get built during our Imagine 2022 Seeed Studio workshop here.

Arducam Pico4ML

The Pico4ML is completely open source and ready to run models built with Tensorflow Lite for Microcontrollers. The onboard display makes it simple to test vision models by showing camera frames along with interference results. With all of the design files available, you can even create your own version of the board that perfectly suits your needs.

  • Cost: $25.99
  • Processing: Raspberry Pi RP2040 microcontroller
  • Memory: 264 KB RAM
  • Storage: 2 MB onboard QSPI Flash
  • Cameras: 1 onboard, 320 x 240 pixel resolution
  • Connectivity: A Bluetooth Low Energy edition is available
  • Power Consumption: ~0.3 watts
  • Size: 51 mm x 21 mm
  • Bonus Feature: 0.96 inch LCD SPI display
  • Edge Impulse: The Arducam Pico4ML is not an “officially” supported board, but thanks to the community, it is documented and works great.

Vision Statement

There you have it, a summary of every single computer vision board currently available*. Take a few seconds to bookmark this page so you know where to find it when you are assembling the bill of materials for your next project.

*Except for all of the other ones we did not cover. Maybe next time around.

Nick Bild
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles