We were initially unsure that we could achieve high accuracy with overhead cameras only. It's definitely not immediately obvious that it should be feasible, and

We were initially unsure that we could achieve high accuracy with overhead cameras only. It's definitely not immediately obvious that it should be feasible, and we weren't convinced until we arrived at a working solution (and a working store!). Getting our action recognition models to high enough accuracy was a big part of the crux, which, not coincidentally, is an area where the literature isn't super mature yet. Our system still makes mistakes, of course, but we're in the high 90's of accuracy, which is good enough economically to start deploying and getting real world users. And there's plenty of signal to squeeze out still.