From a starting point of 86 million photos, the researchers say they were able to find 20,672 time-lapse sequences. The team can cluster images by landmarks and other heavily photographed places, and then sort those clusters by date, the researchers explain in a paper accompanying the project. Each photo is then “warped” so that it shows the same viewpoint, and the entire sequenced is stabilized to make each image progress to the next one with as much visual consistency as possible.
As you’d expect, researchers found more time lapses in regions of the world that are better represented and more heavily photographed on the sites mined by the team. Europe has the highest density of time lapses identified by the system, while South America and Africa had the lowest density.
Each time lapse contains about 1,000 photos, the researchers explained, and takes about six hours for a single machine to complete. “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,” they wrote.
[In Sight: Images of Our World: The Washington Post’s photo blog]
Not every single time lapse identified worked perfectly, or was even interesting, the researchers said. But there were plenty of gems among their haul.
“Some places have quite a seasonal rhythm,” Ricardo Martin Brualla said in a video explaining the project. One example: a time lapse of San Francisco’s winding Lombard Street. The team also observed geological changes at a Yellowstone hot spring and the construction of the Goldman Sachs Tower.
“Waterfalls are particularly interesting, as the streams emerge and dry up,” Brualla noted.
Some of the time lapses are notable for their stillness. Here’s one of the Swiss Guard at the Vatican in which the ever-present guard appears almost like a painting:
The researchers behind this new time-lapse mining project hope their work will only become even more interesting over time. “As more photos become available online,” they said, “mined time-lapses will visualize even longer time periods, showing more drastic changes.”
The team plans on releasing the code for the project, along with more results, in the near future.