A New Algorithm Cuts Down Boring Videos So You Only See the Best Bits

There's one area outside of consumer videos where chopping a video down could be particularly useful: surveillance.

Jun 25 2014, 5:22pm
Image: Refat/Shutterstock

We might have more video than ever before at our fingertips, but that doesn’t mean we’ve got the patience or, indeed, the hours in the day, to watch it all. How often have you recorded events with your smart devices only for the resulting footage to end up forgotten in the growing slag pile of big data? Or perhaps more to the point, how often have you not watched someone else's videos?

Even when it’s cool stuff like GoPros strapped on animals, badass drone footage, or, um Glass porn, there are only so many minutes our YouTube generation is willing to sit through. And let’s face it, most of the video minutes out there are pretty tedious. 

With this in mind, computer scientists at Carnegie Mellon University have developed an algorithm that sifts the interesting bits out of the tedium so you can watch the good bits without sitting through the filler—like a trailer of highlights for boring home movies. If there’s already an algorithm that’ll direct your band’s next music video, there’s now one that’ll edit it for you.

The idea behind “LiveLight” is pretty simple. In their paper, researchers Eric Xing and Bin Zhao explain that the algorithm goes through the video in sections of 50 frames each, building its own “dictionary” of the content as it goes. Basically, it compares a section to the ones that came before it and if it’s too similar, it discards it and moves on. In experiments on surveillance footage and YouTube videos, the researchers found that it took less than twice the length of the original video to create a summary, and in some cases it was able to do so in “near real time.” 

What you end up with is something like this, where amateur footage of a kid being cute but boring is cut down to about half the length. The transition isn’t that smooth—a simple fade—but it gets the main points of interest across.

At the same time, however, the idea of automatically chopping down your device-filmed movies into more palatable bitesize chunks does seem a little clinical, and a roundabout way of addressing the problem of snooze-inducing consumer content. If you don’t have the inclination to do your own edit, why not just make a quick GIF or a Vine in the first place? Or if you’re so sure everyone's going to get bored by yet another video of your kid doing something not at all extraordinary, why document it at all?

That said, the researchers suggested one area outside of consumer videos where chopping a video down could be particularly useful: surveillance. If you’ve got a surveillance or traffic camera running all the time, it’s going to result in a lot of footage, most of which will probably be completely still. Even watching it sped up could take a long time. But as this algorithm would only pick up changes in the scene, it could easily spot moments of interest. The endless frames of an empty street could be discarded; the moment the bank robber swings into view could be retained.

We’ve seen algorithms’ superior ability to analyse surveillance footage before, such as in this one from MIT that tracks human behaviour (and potentially predict it). It's another tool for humans in the security industry, who are already starting to find themselves working with robot colleagues. 

More broadly, being able to summarise video quickly and automatically is an interesting concept in the wider realm of video content. As video has proliferated on the web, it still presents difficulties you don’t get with text. While you can usually find any piece of text you’re looking for with some Google sleuthing and good old cmd+f, there’s no real mainstream way to search video content for the exact bit you want to see. 

LiveLight wouldn’t exactly solve that problem, but, used as a tool to summarise full-sized videos, it could cut down on those moments you want to pull your eyes out while you’re looking for that one climactic scene you could have sworn happened around the 47-minute mark. Film students everywhere, rejoice.

And, I suppose, if you must fill my social feed with your smartphone videos (curse you, autoplay!), at least you could keep them snappy.