I’ve recently returned from a week at #drinktec 2017 in Munich where the latest advances in every field of the beverage industry were on show. VR headsets to show you your new production line, AR goggle so the barman can see the exact status off all his beers on tap and consumer apps the allow the customer to see the temperature of their beer (see blog from my colleague Rosie).
At the Cambridge Consultants stand we were showing two different demonstrations, Smarter Recycling and Vinfusion our wine blending demonstration. On the face of it there is little similarity between the two: serving wine and correctly disposing of your plastic wine glass. But dig a little deeper and they do have something in common, both are sporting cameras, and both are connected systems.
Dig a little deeper still and behind those cameras you’ll find machine vision and machine learning. But whilst both sport machine vision they are chalk and cheese when it comes down to the detail, and of course, the devil is always in the detail.
Vinfusion uses a standard Microsoft Face API to survey the reaction of a customer to drinking their glass of wine. It will give an estimate of the person’s age, gender and their emotional response, it’s an example of being able to capture consumer insights without asking a single question. Deep in the bowels of the cloud Microsoft have trained their machine vision with vast numbers of faces of all ages and races, and have taught the system, trained the system with truth data. Now all we as users of the API have to do is submit and image and back comes the answers. From our point of view all magic as we’re never involved in the training.
Cats and Dogs
Now let’s look at Smarter Recycling, this sports two cameras which look at an object that you are disposing of. Place the object in view and the system illuminates which receptacle it should be placed in to recycle: PET, PP, Compostible or Other. Wow, more magic, it ‘knows’ which materials go where! Actually no, just more training. When we arrived at drinktec we fired up the system and trained it to recognise the items we had on the stand, coffee cups, water bottles, plastic glasses, food containers. Each item it presented in different orientations and positions and the system told “PET”, or “Compostible”. And after a few minutes the system can distinguish between the objects and indicate the correct material type.
The press like to call these systems AI, and without getting into a big debate as to whether this is really AI, think of this comparison. You would teach a child to recognise cats and dogs by showing and saying the name, eventually they understand and will make the distinction themselves, however, still no magic – show them a horse and they would probably think it was a big dog. That is until you taught them, no, this is a horse.
We were asked several times if it used some clever laser surface analysis, the answer is simply, no, it’s working just like you and I, it looks, it recognises the shape and colours and makes a decision. The system does not ‘know’ that a bottle is PET, we told it.
This demonstration shows what is possible, but further detection could be added, identifying brands from logos, even spotting the actual recycling symbol and material type. Fuse all these together and you’re getting even more human in your approach – evaluating multiple data streams.
Horses for Courses
So there are many ways of implementing machine vision and machine learning and it’s important that you evaluate exactly what you’re trying to achieve. Every project is different and may vary from off-the-shelf API to custom hardware and software. As I say: Horses for Courses – provided your trained it to recognise a horse of course.