For many years computer vision was dominated by hand-crafted systems. In recent years a lot of effort was put into generating so-called ground truth data – that is data where the system output is known. To achieve this people have invited interactive games; build complex motion capture systems; Amazon has launched a so-called mechanical Turk labelling system; and many engines we written to synthetize data. This large amount of data had a big impact in our field and data-driven matching learning techniques define these days the state of the art for various application domains. After giving a brief history of this trend, I will present some ongoing research projects in that space, and summarize my thoughts on how to make the next step forward.