Man, software engineering at the cutting edge is getting harder and harder by the day. Not only are you expected to master coding but also math heavy ML and economics. I guess a consequence of software eating the world is that more and more fields of human knowledge is folded into the engineering world. The consequence being that engineers really aught to be very broad in their reading if they want to take advantage of all the low hanging fruit the octopus arms stumbles into.
This is undoubtedly incredibly impressive as a feat of engineering. Detecting codes on those sample images with >95% accuracy is no mean feat. No doubt Google uses its muscle for many worthier things than this, and the learning from this project will be applied to many more 'serious' problems. I think what this blog post triggers for some people is just a further step down the road from Google as it was - a tech first solutions later company that solved interesting problems because they were interesting - to Google of today/the future - a big corporation that acts pretty much like any other, but with a few neat tools in its toolbox.
This is where the current state of the art in industry is in machine learning : do "modern" things like scanning documents/codes/... on old processes without modifying the processes.<p>Say you want to do the famous quality check in factories (is the dogfood box closed properly ?), you can just do that with a convnet.
Really nice.<p>tl;dr:<p>- Customised text recognition of codes printed on goods<p>- UX flow designed to (i) help users correct errors, and (ii) gather additional labelled images