Slashdot Mirror
Breaking Motion Capture Out of the Studio
Fnord666 writes with a CMU press release. From the article:
"Traditional motion capture techniques use cameras to meticulously record the movements of actors inside studios, enabling those movements to be translated into digital models. But by turning the cameras around — mounting almost two dozen, outward-facing cameras on the actors themselves — scientists at Disney Research Pittsburgh, and Carnegie Mellon University have shown that motion capture can occur almost anywhere — in natural environments, over large areas, and outdoors."
Interesting approach. I would have loved to have seen the presentation, but also today at SIGGRAPH, Autodesk had a presentation by Kevin Wittkopf and Geoff Richardson showing off using the Kinect in production with Maya. Why bother with a suit and a bunch of cameras when Kinect rigs and only going to be getting better (and waaaayyy cheaper) Autodesk is streaming their presentations from the exhibit hall. I'm not sure if they are going to do that prese again tomorrow or Thursday, though. (please don't everyone check it out at the same time...I'm trying to watch it)
Just wait for version 2.0 where the cameras are smaller than the size of a coin and everything is neatly sewn into a spandex motion capture suit. Think of potential over real terrain, rock climbing, tree climbing, etc.
So how long before the inevitable application of "Rule 34" to this?
Let's face it, the porn industry has been a leader in adopting technologies for some time now.
Lost at C:>. Found at C.
Well, it's mathematically complex, but not really complex in a code size or if-then count sort of way.
Here are two videos (1, 2) from 2007 showing the jist of it (including some games that modify reality) done in *realtime* using a laptop GPU card.
It's basically...find a bunch of "features" in a scene (corners in some texture, circles, etc), then look at how these features move in the scene as the camera moves. Each feature is in a fixed point in space, so, their movements are caused by the camera moving. Using basic multi-view type math from the multiple views brought on by the movement and each frame of the video, you can "solve" the cameras position, rotation, AND all of the feature positions in 3D space using a big giant matrix created by all of those 2d image feature positions. If you've taken first year of linear algebra, and you covered singular value decomposition, you'll understand probably 80% of the math in the research papers (they'll find crazy efficient ways to do things...though.
Some complexity comes from disregarding moving objects in your scene...but that's mostly a "throw out data with a wtf factor greater than x" type statistics problem.
GPU's are very good at finding features, which is why that video from way back in 2007 is possible...that's around the time that shader programs with reasonable complexity could be made (like more than three texture lookups per pixel). The lack of complexity is also proven by the huge number of flash applications a few years ago that used this with your webcam, and the existence of the Nintendo DS AVR games, and all the cellphone apps that use it.
This isn't bleeding edge or all that complex...but it is a very cool. It's more of the standard, "the hardware and software is cheap enough to do it now".
If you're interested in this stuff, you could get something reasonable going using OpenCV pretty quickly...most of the feature tracking routines are built in.
On a side note...look at this cool shit we could be working on if people would stop messing with f'n web browsers/apps...