10x Management, the leading resource for companies looking to hire world-class freelance tech developers, data scientists and designers is pleased to announce that Software Developer Greg Sadetsky and his team member Ann-Julie Rhéaume have been awarded second place in the U.S. National Library of Medicine’s Pill Image Recognition Challenge.
The Pill Image Recognition Challenge launched in January 2016, and called for submissions from teams to contribute to the creation of a software system that could match photos taken by a smartphone to the NLM RxIMAGE database of high-resolution prescription pill images. This would give consumers a simple way to recognize mystery pills, help prevent unnecessary medication errors, and reduce waste by identifying pills that might otherwise be discarded. The winning teams were announced on August 1, 2016.
Greg states, “Ann-Julie and I first started by splitting the work: she was responsible for the machine learning using Google’s TensorFlow, while I was responsible for the image processing using OpenCV (locating the pill in the photo and getting rid of the photo’s background) and for the creation of the virtual image that we sent in which involved setting up the operating system, installing and compiling the required libraries, etc. The basic machine learning and image processing split allowed us to work independently. We both were doing everything we could to increase the matching accuracy of both of our respective sides — I would try to match pills shown on different backgrounds and in different lighting environments, while Ann-Julie would work on the neural network to increase its ability to match reference pills to the images taken by everyday consumers.”
When asked what the hardest or easiest part of the challenge Greg replied, “Like any research and development work, what’s exciting is also what makes it hard. You don’t know what’s going to work, so you: A) explore a few options (different algorithms) then B) try to combine those algorithms and see which combination seems to be the most promising and then C) tweak the promising combination of algorithms to improve the accuracy score. Rinse and repeat, start from scratch many times, keep many different research directions open at once, etc. Creating the virtual image was pretty easy; I took a little bit of time to automate the installation and compilation steps, which really helped. Deploying in non-automated ways means that you need to remember a lot of ‘magic tricks’ applied once, and when the pressure rises, those are usually the first steps that you will forget to apply, creating an avalanche of problems.”
And there were some challenges along the way. “Creating image processing pipelines is difficult. There’s a lot of specific domain knowledge and ‘tricks’ that you learn with time, but it’s still surprisingly hard. Our brains are good at fooling us into thinking that because we, as humans, can see and identify objects naturally, it “should not be that hard” to teach the computer to do the same. Well, it truly is hard. That’s humbling,” said Greg.
“We are so proud and honored to have Greg as one of our 10x clients. The dedication, world-class talent and impactful outcomes that 10xer’s like Greg Sadetsky demonstrate on a daily basis is exceptional. His talent is usually bestowed on our customers but in this instance, it was put to use for the greater good. I have needed to use pill identifiers in the past for various needs in my family and the problem they solved was a very real one. Plus, he’s a hell of a nice guy,” says Michael Solomon, Cofounder of 10x Management.
From all of us at 10x Management congratulations Greg and Ann-Julie!
About 10x Management
10x Management is the world’s first tech talent agency, matching innovative companies with the best freelance developers, designers, and coders in the world. With offices in New York and San Francisco, 10x works with leading companies including: Genentech, eBay, Yelp, BMW, American Express, Verizon as well as hundreds of venture backed start-ups. Inquiries may be sent to: firstname.lastname@example.org
About Greg Sadetsky