Whenever I see something like this I'm reminded of the introduction to "Multiple View Geometry in Computer Vision", which is where a lot of this "clever math" is described. Essentially, problems considered unsolvable ten years ago are not only solvable, but simple, due to advances in geometry. With just a little math, you can do things like this. I wonder why it took this long for the people doing computer vision work to catch up with the math.
Because most commercial applications want accuracy and scale.
So use structured light, background target grids or multiple fixed cameras.
This technique needs fairly simple shapes and assumes that edges are square etc. You couldn't use it to scan a clay model for CGI or a complex shape for cad. You can use it to convert simple 3D objects into CAD but that's a bit pointless compared to just drawing them in CAD.
It's clever, congratulations to Mr Pan, but it's not as earth shattering as the press releases claim.
While there have been countless important geometric and mathematical insights applied to/developed in computer vision in the last decade, it's really the rapidly increasing computational power available on commodity hardware that has made many advances in real time vision possible.
During the four years I was doing my PhD [in real time vision, a couple years ahead of Qi Pan in the same group], the amount of useful image processing and linear algebra/geometry that could be computed in 33 ms jumped dramatically.
The computer vision community is collectively realizing that techniques that were always mathematically and statistically sound, but considered too expensive for online processing, can now feasibly be incorporated into real time and interactive systems.
There are fewer rather than more 'tricks' now, and more straightforward representations of the underlying mathematical and probabilistic structures. This is a good thing.
Likewise, real-time reconstruction techniques appropriate for desktops in 2004 can now run very well on embedded platforms.
It's an exciting time for computer vision, as the set of feasible applications is exploding.
Does anyone have a working open source implementation of this?
I saw this video about a week or two ago, and decided to hook up at webcam to python / pygame and try a few things, and I have trouble even checking what colors changed each frame in under 1 second, I wonder what methods they are using.
The open CV lib is in C (with IPP cpu-optomized assembler if you want) calling the functions from python doesn't slow you down - unless you do something really dumb like trying to itrerate over pixels in the python cde.
