My bet is they just invested a few months making a varied dataset, a few grand on mturk using labelme, some image augmentation, a few engineering tricks for a nice UX,some strategies for getting the model to converge on tricky test images, and got it to hit 100% accuracy fairly quickly.

We are working on this product in-house for 7 years now. We are an internal startup from a company that does custom computer vision and ML software.

No mturk, no tricks, a lot of work in ML and in UX as it is not a very straightforward task.

> It is not a very straightfoward task.

I'd well believe it. I used to write computer vision applications for semiconductor manufacturing equipment and there we were able to strictly control the distance from camera to object, lighting etc. and even still getting necessary reliability was not simple. When a failure could lead to damaging a whole wafer, i.e. hundreds of thousands of dollars, 99% accurate is not good enough.

It would be very questionable to use this in production, especially as pearls are very glossy and costly, without the error matrix being provided. It would seem more reliable to use the camera as a blocked / non-blocked sensor with what would amount to a cheap coin sorter.

For some reason it's very common here on HN to trivialize the difficulty of automating tasks. But you're not new here so you probably are familiar with it!

Wasn't trying to trivialise your work! Just offering some ideas of the basic ways this could be tackled. Sorry if you thought I was trivialising it, the intent was more to support the idea that modern DL can still provide solutions with relatively simple methods.