First-person video cameras like the GoPro range, Garmin’s Virb and Drift’s HD Ghost are growing in popularity, but getting footage from them that doesn’t induce nausea still requires at least a modicum of video-editing skill. A new “hyperlapse” project from a three-man team at Microsoft Research wants to change that.
There are two immediate challenges when trying to turn first-person footage into something visually arresting – the first is sifting through and trimming the (often lengthy) footage down to the attention-worthy parts, which many people get around by using every tenth or twentieth frame to create a timelapse video, which compresses the footage substantially.
Another way around this problem is to edit the footage so that only the most interesting events are included, but this leaves the user with gaps that need to filled with transitions or other editing techniques.
The second challenge is the jerkiness introduced by things like bumpy roads, wind or, in the case of helmet cams, looking around while filming. When going the timelapse route, jerkiness is often exacerbated, resulting in footage that’s taxing to watch and unlikely to win any YouTube views from anyone other than friends and family.
Microsoft’s hyperlapse project – a hyperlapse is a timelapse video that gives the appearance of having been short with a smoothly moving camera – seeks to alleviate these problems by using some intensive algorithmic trickery and a bunch of processing power.
The software takes the existing footage and creates a virtual, 3D path from it. It then reworks the footage to ensure that each frame follows this virtual path consistently – automatically and intelligently filling in any gaps as it goes. Finally, the software generates a smooth timelapse video by rendering, stitching, and blending the applicable source frames for each subsequent output frame.
A more technical explanation follows in the video below. The team behind the project is hoping to release the algorithm as a Windows application in due course.