The logo detector developed in this chapter does not continuously update its network weights after being deployed. Once it is trained to recognize a set of logos, it should continue to do so with the same accuracy. There is a chance that, after being deployed, the kinds of photos people take gradually changes over time. For example, the increased popularity of Instagram filters and related image manipulations might begin to confuse the logo detector.
In any case, it is still important to continuously evaluate whether the detector is working as expected. This is a somewhat challenging exercise because it requires that humans are in the loop. Every photo with a detected logo can be saved to a database for later examination. Our logo detector code does this. Every so often, a team of people can be asked to critique the tool's predictions to produce a continuously updated measure of precision.
Measuring recall is more challenging. Among the universe of photos shared on social...