Real World AI: Warehouse Recognition System

Solution that immediately comes to mind is building a system that’s driven by Computer Vision and tells you the name of the item of interest and shows the location in a warehouse where similar items are stored.
Direct Solution is to take tens of photos of each part (out of about 10 thousand) and use those to train a parts classifier. And then as new parts will continue being added into the catalog take more photos and retrain the system. Solution would work, but... creation of such a training dataset would take several months of labor and thereafter will need permanent control over the growing dataset. But Customers usually want to get results faster, cheaper and without necessity to regularly pay for expensive manual retraining of the system.

Is there anything we can do about it?

We were lucky and the answer in our case was yes. A nuance that allowed us to significantly reduce the cost and duration of the project is that in our case we deal with a factory warehouse, and stored parts are all flat, although they are made from steel sheets of different thicknesses. And more importantly, there are CAD models for all the parts.

Learn more about the story on Medium.