A team at Stanford is using Google Street View images to successfully predict election results. Aside from the sheer vastness of information publicly available on Google, the big takeaway is that a preponderance of sedans indicates that a given area is likely to vote democratic, while pickup trucks indicate republican leanings. And the imminent obsolescence of human polling data.
The researchers started with a deep learning AI algorithm designed to be good with visual inputs, and trained it to recognize not just the difference between cars and trucks, but to break them down by make, model and year. That might sound complicated, but it’s just scraping the surface of the project. Once the AI tool—also known as a convolutional neural network, or CNN (yes, really)—was ready, the team had it analyze 50 million street view images.
Google Street View Image, Lubbock, Texas
In just two weeks, the neural network scanned and classified over 22 million vehicles across 200 cities, which equates to roughly eight percent of all vehicles on U.S. roads. To put that into perspective, a human looking at each image for just 10 seconds would require 15 years to identify that many cars and trucks.
Perhaps even more impressive, the CNN doesn’t just determine if it’s looking at a car or truck based on shape, analyses minute details that even all but the most enthusiastic automotive experts wouldn’t notice. For example, it can recognize the difference between a 2007 and a 2008 Honda Accord based on a subtle difference in the taillights.
By comparing the wide-scale data with known election results and U.S. census data, the team was then able to note clear correlations at even the most granular levels. If sedans outnumber pickup trucks, there is an 88 percent chance a precinct votes Democratic, whereas if pickup trucks outnumber sedans, that chance becomes 82 percent in favor of Republican.
While the researchers caution that the data is not a substitute for a comprehensive census poll, they do point out that it is an incredible tool and could replace polling in poorer countries that can’t afford such an expensive undertaking.
The true value of this new source of data is hard to fathom. Because general price of each vehicle can be easily estimated, the data will of course be highly prized by insurance companies and financial institutions. Vehicle type and crime rate data is also out there, waiting to be discovered.
Ultimately, the possibilities for uses of this data extend far beyond elections.