Machine vision system in action

Back in November, in a post entitled, “NeuroCheck machine vision application.” I introduced you to a system that sorts the little plastic size tags that you see on clothes hangers in the store. Well I’ve just received an email from the company that developed the system, Industrial Vision Systems Ltd., with a link to a YouTube video.

To save you the trouble of searching YouTube, I’m embedding the video below.





In my November post I commented that the reported accuracy of the system was less than 100%. Earl Yardley, Director of Industrial Vision Systems explained that the errors, or misidentifications, are generally due to dirt or damage of the individual tags, in which case it might be argued that the system is 100% accurate.

This brings up what I consider a vital yet neglected issue: just how should we define the performance of a vision system? The problem I see is that while the system developer will say that a rejection due to dirt means the system is operating as intended, the customer will often say that it is not doing what he expected it to do. It’s all a question of perception or point of view, but it’s also an important question because it’s at the heart of so many buyer-seller disputes over performance, and of course, payment.

What do you think? Is this an issue that you have encountered, and do you have any advice to offer?