In many ways, the internet is simply an extremely badly coordinated attempt to document our planet. Hundreds of millions of photos are uploaded online every day, depicting ourselves, our cities, our landscape, and our food. However, they're also isolated and disparate: it's difficult to sieve anything meaningful from the avalanche. Researchers from Google and the University of Washington have found a way though, creating a powerful set of algorithms that automatically sorts millions of online photos into time-lapses of everything from skyscrapers to glaciers. The internet is plugged in at one end, and a record of our changing world comes out the other. They call it "time-lapse mining."
"We can now almost instantly create thousands of time-lapses."
"Ultra-slow changes are documented by the billions of photos that people take over time," write Ricardo Martin-Brualla, David Gallup, and Steven M. Seitz in a paper documenting their work. "Whereas before it took months or years to create one such time-lapse, [with our algorithms] we can now almost instantly create thousands of time-lapses covering the most popular places on earth."
They also reveal changes in the natural world, such as the retreat of the Briksdalsbreen Glacier in Norway. (University of Washington/Creative Commons)
This time-lapse shows the snake-like movement of a sandbank in Thailand over time. (University of Washington/Creative Commons)
One of the chief difficulties for the team was synthesizing a single focal point for their time-lapses. Fortunately, the sheer number of images available and our tendency to photograph the same landmarks from the same angles gave them a good starting point. They created a database of some 86 million publicly available photos from sites like Flickr and Picasa and began sorting them; first, creating clusters of different locations and then putting these sets into chronological order. For each time-lapse they selected a single reference image that would form the anchor point for the other photos, with their algorithms cropping, aligning, and color-fixing the images to create a single, smooth time-lapse.
A time-lapse of 1,000 crowd-sourced photos took six hours to process
The team says that in the end they were left with 120,000 clusters, with 10,728 of these containing more than 300 images. From these candidates they selected clusters with more than 1,000 photos to process. "A typical time-lapse with 1,000 input posed photos takes about six hours to compute on a single machine," report the engineers. "While the algorithms can be optimized a lot more for efficiency, we point out that a few hours is negligible compared to the time period of several years it took to capture the photos." The end results are crowd-sourced time-lapses of the changing world, documented almost accidentally by millions of volunteers.