Stabilizing Video

A new system cleans up shaky amateur footage.






Steady hand: New software from Google focuses on key elements of a video to stabilize footage. Credit: Matthias Grundmann
Source: "Auto-Directed Video Stabilization with Robust L1 Optimal Camera Paths"

Matthias Grundmann et al.

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, Colorado, June 21-23, 2011

Results: Google researchers have developed a technique that removes the effects of an unsteady filming hand on amateur video. After showing that they were able to smooth out footage in real time while maintaining focus on key elements in the image, they used their algorithm as the basis for a stabilizer application that runs in real time at www.youtube.com/editor. YouTube, the largest site used for sharing amateur video, did not previously have a stabilizing feature in its nine-month-old Web-based video editor.

Why it matters: Shaky footage has always been a mark of amateur filmmaking, since amateurs usually lack the costly equipment that professionals use to stabilize their cameras. Some algorithms are available to clean up this effect, but those typically remove jerking effects without correcting others, such as the slow bounce of a camera held by a person walking.

Methods: The algorithm begins by identifying key objects in the image and using them to plot the path traveled by the camera. It then determines a "best path," which represents the smoothest course for the camera to have traveled. By cropping the frames, it is able to adjust the footage so that the camera appears to have traveled the best path. The algorithm uses tools such as face detection to make sure it doesn't remove key components of the video in the process. Because the computational work is distributed among many machines, the system is fast enough for editing in a browser in real time.

Next Steps: Right now, the user has to specify what size the system should crop the frames to; the algorithms find the optimal path for the given size. In the future, the researchers plan to adjust the system so that it can calculate the ideal size on its own.