Tagged Photos are Perfect for Surveillance
For decades, researchers have refined the ability of computers to learn from acquired datasets and use that information to perform tasks. The worry has been that these techniques would lead to overly good surveillance systems.
In general, they have not. The biggest bottleneck has been the lack of good “training” datasets to feed to the computer. In order for a computer to learn how to find a person in an arbitrary video sequence, it needs to be fed a pre-exisitng body of images or videos of that person in a variety of poses, lighting, and other variations.
It is not enough to just hand the computer photos of a person. The computer needs to know where in the photo the person actually exists to disambiguate between other people that may be in the photo and other background objects which may be hard for the computer to automatically label.
Tagged photos are the perfect source of data for feeding machine learning algorithms for surveillance systems. The tagging identifies the person and where in the photo the person exists. Enough tagged photos of the person can provide the computer with the information it needs to attempt a reasonably accurate classification using future video or photo streams.
Tagging photos can happen through semi-automatic computer vision algorithms. But a way to get large sets of tagged photos is through crowdsourcing, e.g. on Facebook.
Tagged photos are controlled by only a handful of companies and represent incredibly valuable crowd-generated information. If Facebook launched a massive machine learning driven surveillance effort at the upcoming NCAA tournament, I wonder how well the system would perform (i.e. accurate classifications vs inaccurate classifications vs indeterminate).
What are your thoughts on how tagged photos change the game for surveillance systems?