I worked in the object recognition and computer vision industry for almost a decade. That stuff works. Really well, actually.

But this checkout thing from Amazon always struck me as odd. It's the same issue as these "take a photo of your fridge and the system will tell you what you can cook". It doesn't work well because items can be hidden in the back.

The biggest challenge in computer vision is occlusion, followed by resolution (in the context of surveillance cameras, you're lucky to get 200x200 for smaller objects). They would have had a really hard, if not impossible, time getting clear shots of everything.

My gut instinct tells me that they had intended to build a huge training set over time using this real-world setup and hope that the sheer amount of training data could help overcome at least some of the issues with occlusion.