Time-lapse Mining from Internet Photos
"Synthesizing a photorealistic time-lapse is also very challenging," explains Martin, "as the input photos look very different from each other ... Photo timestamps might be wrong and many photos contain occluders, like people posing in front of monuments, that our time-lapse successfully ignores."
It's a complicated business, but also fully automated: Just feed the algorithm, and in a few hours it'll spit a video right back for you. Their efforts yielded nearly 11,000 time-lapses, most of which seem to be cover temporal stretches of five and 10 years. Among the highlights are the erosion of the Briksdalsbreen Glacier in Norway, the rise of New York City's Goldman Sachs Tower, and a Swiss Guard at the Vatican who remains still enough over six years that he becomes every bit as much a part of the time-lapse as the iron door frame around him.
In a random sample of 500 time-lapses, the researchers found nearly half to be "good and interesting," which is to say they have no visible artifacts, are photorealistic, and actually show something changing over time; Martin points out that indoor scenes will often simply look like still photographs, since the changes even over a span of years can be too minute to notice.
The team hopes to be able to automatically filter out the boring results. In the meantime, some of those "failures" are captivating in their own right. The position of the famous Wall Street Bull has actually moved slightly over the years, creating a blur effect and a quick reexamination of your assumptions about statues and mobility.
Previously, previously, previously, previously, previously, previously, previously.