A short Two Minute Papers YouTube explainer (Aug 25, 2015) describing methods for creating high-quality time-lapse videos from large collections of community photos. The speaker frames the problem (simple mining of hundreds of Internet photos produces flicker and artifacts) and lists concrete causes — differing illumination, season, viewpoints, and per-image focus differences — calling the alignment/normalization task “regularization.” The video highlights improved results from a recent University of Washington and Google collaboration that dramatically reduces flicker and reveals slow changes (examples: Lombard Street seasonal rhythms, sculptures, and a Norwegian glacier timelapse), and contrasts these with earlier algorithm outputs.
Chapters
Hey there, fellow scholars. I amai and this is two Minute papers where we learn that research is not only for experts, it is for everyone. Is everything going fine? I hope you are all doing well and you're having a wonderful time. In this episode, we are going to look at time lapse videos. Let's say we'd like to build a beautiful time lapse of a Norwegian glacier. The solution sounds quite simple. Just mine hundreds of photos from the Internet and build a timelapse video from them. If we just cut a video from them, where we put them one after each other, we will see a disturbing flickering effect. Why? Because the images were taken at a different time of the day. So the illumination of the landscape is looking very different on all of them. They are also taken at a different time of the year and from different viewpoints. Moreover, since these images are taken by cameras, different regions of the image may be in focus and out of focus. The algorithm therefore would have to somehow equalize all of the differences between these images and bring them to a common denominator. This process we call regularization, and it is a really difficult problem. On the left you can see the flickering effect from the output of a previous algorithm that was already pretty good at regularization, but it still has quite a bit of flickering. Here on the right, you see the most recent results from the University of Washington and Google compared to this previous work. The new algorithm is also able to show us these beautiful rhythmical seasonal changes in Lombard Street, San Francisco. It can also show us how sculptures change over the years. And I feel that this example really shows the possibilities of the algorithm. We can observe effects around us that we would normally not notice in our everyday life, simply because of the reason that they happen too slowly. And now here's the final time lapse for the glacier that we were looking for. So building high quality time lapse videos from an arbitrary set of photographs is unbelievably difficult. And these guys have just nailed it. I'm loving this piece of work. And what do you think? Did you also like the results? Let me know in the comment section. And for now, thanks for watching and I'll see you next time.
