The world’s leading publication for data science, AI, and ML professionals.

Being better at Machine Learning than Google – is it possible?

To say you're better at something than Google is fight'n words. However, when I ran a test of our facial recognition technology against…

Comparing Facebox by Machine Box with Google Vision
Comparing Facebox by Machine Box with Google Vision

To say you’re better at something than Google is fight’n words. However, when I ran a test of our facial recognition technology against Google’s Vision API, I discovered that we were more accurate… a lot more accurate.

When I saw ‘we’, I’m talking about my company Machine Box. We make machine learning models inside Docker containers so that any developer who can write to an API can implement machine learning in their products, no PhD or math required.

This all started out as an experiment to see how many faces Facebox can see in a single image. I searched for pictures of ‘crowds’ to find some good test images and came across this:

A photo with a lot of faces - source: http://www.sportsplexoperators.com/blog/wp-content/uploads/2013/07/crowd_2.jpg
A photo with a lot of faces – source: http://www.sportsplexoperators.com/blog/wp-content/uploads/2013/07/crowd_2.jpg

So I spun up an instance of Facebox and posted the picture to it. After a few seconds of downloading and analyzing the image, Facebox responded with the JSON, at the top of which is a facesCount element with the total number of faces detected.

JSON response from Facebox
JSON response from Facebox

153, not bad. I ran the picture again through our developer’s consoles preview mode to get something worthy of a screenshot:

Screen shot of Facebox's preview mode in the developer console
Screen shot of Facebox’s preview mode in the developer console

I then wondered how this stacked up with other cloud cognitive services, particularly Google Vision, in light of their new announcements around Machine Learning. I visited their page where they let you post a photo to the service to see what the results would be. After a few seconds of processing, I got this as a result.

Google Vision console
Google Vision console

I could immediately tell from the thumbnail that fewer faces were detected, but when I scrolled down the response window on the right side I was a bit shocked to see that they had only detected 50 faces.

They missed over 100 faces in that photo. Thats a lot. 😬

Google Vision certainly has a lot more features in their face recognition product with things like emotion detection, eyebrow positioning, headwear detection and more. At Machine Box, we made Facebox to be really good at detecting and recognizing faces (you can teach Facebox who someone is with a single image). That’s it! We do this because we want our boxes to be of the highest possible quality, which means only implementing features if the state of the art is good enough to return accurate results.

That being said, we love Google, and for some use cases, their public cloud APIs for machine learning will work great. They have an amazing data set with which to train machine learning models. So is Machine Box better at machine learning than Google? No, of course not… its Google! They’re awesome for investing a lot of time and money into deep learning and other AI-related technologies, and of course, providing valuable services to just about everyone on the planet. But a lot of developers out there need to implement something on premises, or in their own infrastructure, that has an extremely high degree of accuracy (and they love Docker), so they go with Machine Box – and we couldn’t be more grateful.

Making razor sharp developer tools is our passion and creed, and we’d love for you to give it a try. It will always be free for developing and testing.


Related Articles