I compared my startup’s face recognition technology against 3 of the big cloud vendors. Results below.

I recently joined a startup called [Machine Box](https://machinebox.io), and one of the first things I did was test our technology against three big cloud vendors. After all, machine learning is only useful if its accurate.

These cloud vendors provide machine learning as a service (sometimes referred to as cognitive services). They usually charge per API call and require sending data (like images) to a public API endpoint.

One of the key value propositions of Machine Box is that it can run on premises. No need for sending data out of your infrastructure. But can Machine Box compare in accuracy to the big guys? I wanted to find out.

To keep things simple, I decided to compare facial recognition technologies. Machine Box’s facial recognition technology is called [Facebox](https://blog.machinebox.io/streaming-and-recognize-people-from-a-webcam-with-go-and-facebox-acea645b94ab). I picked an image with a lot of faces and ran it through [Facebox](https://machinebox.io/docs/facebox) to see how many faces it acurately detected. I then ran the same image through the big cloud vendors’ cognitive services. These were my results:

![](https://hackernoon.com/hn-images/1*aZn2K7eEK2Dj4tHoqY4aHg.jpeg)

Facebox detected more faces than the other big cloud vendors’ machine learning services

Facebox was able to pick up every single face that was not obscured in the photo. Two of the three cloud vendors came close, but missed a few. One vendor didn’t pick up any faces.

Next, I wanted to compare the celebrity recognition capabilities of these technologies. I used [one-shot teaching](https://blog.machinebox.io/machine-learning-model-training-vs-machinebox-teaching-f959f5fa4635) to train Facebox on every US politician, then picked a random photo off the internet of some members of the US House of Representatives.

![](https://hackernoon.com/hn-images/1*a2g_QJRoYYC646rVYlA9HQ.jpeg)

GOP members of the US House of Representatives

I posted the photo to Facebox and the same three cloud vendors. These were the results:

![](https://hackernoon.com/hn-images/1*UFo_XatvCr3U6nDVqMCvAg.jpeg)

Comparing Machine Box’s celebrity recognition to 3 big cloud vendors

Facebox performed even better than expected. Oddly, the cloud vendor that failed to detect any faces in the earlier test was able to eek out a meesley 13 known people in this image.

![](https://hackernoon.com/hn-images/1*0JD72geFRsIPwZ1se3gRsg.jpeg)

Cloud vendor number 2

The first cloud vendor recognized two members of the house. The third cloud vendor doesn’t offer a celebrity recognition service, or any easy way to train it with known people.

![](https://hackernoon.com/hn-images/1*N_H9JZk0ol1VrnlW5xQ5PA.jpeg)

Facebox recognizes almost every member of the House in this photo

Facebox picked up all but a handful of members of the house.

When thinking about facial recognition or celebrity detection, it might be wise to consider Machine Box.

Keep

Too Long; Didn't Read

Get hands-on learning from ML experts on Coursera 

Which face recognition technology performs best?

Too Long; Didn't Read

Companies Mentioned

@aaronedell

RELATED STORIES