I wanted to share a video of some of the application research we have going on at our Intel China Research Center in the area of video mining. In collaboration with Tsinghua University, Yimin Zhang and his team at ICRC are trying to create improved computer vision algorithms for use on future multi-core hardware. The idea is to allow computers to recognize content in professional or user-generated video so that one can automatically edit (or at least pre-edit) to isolate only the most interesting parts of the video.

In this video you’ll see that the target application in this case is extracting video highlights in a soccer game. Many people can relate to having hours of sports video stored up on their DVRs, especially during a playoff season. You might not actually have the time to watch the all the footage (especially if you’ve already heard the outcome), but you might want to see the highlights — just the players and events that interest you most.

The video has three segments. The first shows the basic capability to recognize some of the features in the game — a player, ball, a goal post, etc. The more cores you have, the more complex the features you can recognize in a reasonable amount of time.

The second segment shows a scalability demo — a “bake-off” between a 1-core an 8-core processor. This was done to show that video analysis is truly a parallel application. The graph at the end shows that the more cores you have the better performance you get.

The final segment is a concept application — a future DVR with the capability to pick scenes from a game depending on what happened.

The idea is to be able to do this on any match video — professional or home brewed (in this case some Intel researchers playing around) without having to manually pre-tag the data. It’s one of the capabilities we’re trying to enable through tera-scale research. In fact, Yimin and team will be presenting the latest results on this and related video mining work at the Intel Developer Forum in Shanghai in just a few weeks.