03. Automating the Automation

Artificial Intelligence

How do you count millions of tiny objects accurately and cost-effectively? Artificial intelligence of course.

Having built display stands with remotely managed miniature cameras giving full shelf visibility – how would we go about the task of counting products to monitor true on-shelf availability?  The answer we settled on was training custom-vision AI. 

Over a series of weeks, smart display units collected thousands of images. By using 1,000 reference images to train the AI we could accurately predict every product in the display to 99.6% accuracy.

Having proved the concept with incredible success, the next problem was scaling up the training. Over a hundred man hours had been spent training the AI for a small subset of items, this would never be viable for the hundreds of items that could appear in a display. Time for some fresh ideas.

We needed to train the AI without any imagery.  Working with our development partner and their talented 3D team, we put them to work modelling the packaging of products while the development team set to building a training engine to create highly detailed renders of products and display patterns that could be used to automatically train the AI.

With the ability to create thousands of 3D renders to the same specification of the smart displays we could now generate thousands of images a second that looked just like the real thing.

The result was a configurable training algorithm that can be used and re-trained each time the real-world display is updated.