If you need to add machine learning capabilities to your next project, one option is the Intel Neural Compute Stick 2 (Intel NcsCS2). This device acts as a plug-in AI accelerator and compares most closely with the Google Coral TPU Edge Accelerator. Both devices connect to the host computer via USB ports and are specialized for machine learning.
We'll focus on how to set up the device and get it running for the first time. As noted on the startup page, the Intel Ncs2 can run under:
- Ubuntu Linux
- Raspbian (Raspberry Pi's official Linux distribution)
- Windows 10
The Coral USB accelerator runs on Linux/Raspbian and Macintosh, so your OS capability and preferences may dictate which device is right for you. Notably, the Neural Compute Stick requires an x86-64 computer on which to run, so I didn't attempt it on my aging x86-32 Ubuntu laptop. Instead, I opted to use my other laptop, a Windows 10 machine built in the last two years.
Downloads and Yak Shaving
In addition to the OpenVINO toolkit required for the Intel Ncs2, you'll need several weighty dependencies. For now, however, navigate to the OpenVINO download page and select the OS you'll be using. After entering your email address and other info, your browser should automatically navigate you to the "download" button. The browser didn't work correctly for me when I clicked through, but I used a confirmation email successfully instead. Download your preferred version, which will likely be the latest.
Open the downloaded file and follow the installation instructions, noting any missing dependencies, such as:
- Cmake version 3.4 or higher
- Microsoft Visual Studio 2015, 2017, or 2019
- Python 64-bit version 3.6 or higher
In my case, I already had Python installed, so that wasn't listed.
Once the installation is complete, follow the instructions under item two to install the dependencies software you need. You can choose from several options, but the default selections will probably suffice.
Downloading and installing the files takes time, but I was pleased that everything installed on my system without issue. One thing to remember when installing CMake is to "Add CMake to the system PATH for all users."
Plug in and Run First Example
With the four necessary programs installed, it's time to run your example.
- Plug in the Intel Neural Compute Stick to a USB drive.
- Open up a command prompt, and navigate to the directory C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\demo.
- Run the command: .\demo_squeezenet_download_convert_run.bat –d MYRIAD.
If successful, you'll see a display showing classid, probability, and label, "Sports car, sport car" should be listed in the first row. Congratulations, your device is working.
License Plate Identification
If generating a bunch of text isn't exciting enough for you, you can also visually identify cars and read their license plates.
- Navigate to the same demo directory as before, then input: .\demo_security_barrier_camera.bat -d MYRIAD.
- This command will identify a car in a car_1.bmp file stored in this directory and read its license plate.
- Another window will pop up with the image, and the demo will be complete when you close it.
More Automotive Tests
To test the device's ability a bit further, I navigated to this directory in the program manager, then replaced the car_1.bmp file with another image I found on Wikimedia (converted to a bitmap—BMP format in Photoshop Elements). You'll need to provide administrator permission to change the names around, and I suggest also renaming the original image so you can find it later. Once complete, I reran the program, generating the following image:
I had assumed the device wouldn't struggle with this image, but perhaps it didn't like the straight-on view or the car's classic lines. For whatever reason, it did not recognize this as a car, nor did it see the license plate. Perhaps the demo needs modifications for different angles, car styles, or license plates.
I did throw in one image that I expected it to miss completely: a toy Minnie Mouse car with a "license plate" that I made out of a Post-it note. Interestingly, it did recognize the image as a car, though the NCS2 apparently didn't appreciate my attempt at a very shoddy plate. It would seem the mouse-mobile is safe from surveillance, at least for now.
Conclusion
After getting my hands on an Intel Ncs2, it's a brilliant-looking device that, unlike the Google Coral TPU, runs on Windows. Installation and setup take time, but the process went smoothly enough for me. The demos are also straightforward, though running them on images I provided doesn't seem to work as well as I'd hoped.
Regardless, with some tweaking, it might be just what you need—and seems to be just about the only option in this class of device if you need to work with Windows. In addition to their documentation, the NCS2 has a series of three videos to help you through this process, all of which are available on Intel's "Getting Started." I found the Windows version quite helpful when setting everything up.