How does a computer know a car is a car, even if it’s a different shape, or color etc.
In: Technology
Well, in a way, we don’t know. Google and their competitors all have algorithms that look at pictures and try to analyze them. These algorithms must be trained by being told what is and isn’t a car. They are trained with huge amounts of data.
Google’s captchas are actually part of how they train theirs. They have the user click on all stop signs or streelights or whatever and then their AI learns from this data.
But because we are training it, the program is sort of writing itself, and so we can’t always say exactly why it works.
The other comments touch on how we train a computer to tell cars apart from other objects, but some more info on how it actually works in the end: The most common types of models for object detection have many layers. The early layers detect a bunch of little tiny features in the image, things like lines, corners, blobs, etc., and they find those features anywhere they occur in the image, by checking each little patch of the image. Later layers look at the features detected by the earlier layers, looking for specific arrangements of those features.. for instance, they might recognize a certain arrangement of lines as a wheel. Then the next layer might recognize that two wheels, so far apart, and some windows above that, is probably a car.
So, basically, the model starts by recognizing tiny pieces of things, then uses that to recognize bigger pieces, building up to recognizing entire objects.
Latest Answers