Re: [agi] Real-time Java-based vision libraries?
On 28/01/2008, Benjamin Johnston <[EMAIL PROTECTED]> wrote: > I'm after either: fairly robust object segmentation; 2.5D or 3D > reconstruction; object recognition; or something like superquadric > reconstruction. Something that attempts to describe the physical structure > of the world seen through the camera. These are all non-trivial problems and I don't know of any libraries (java or otherwise) which "out of the box" perform 3D reconstruction in real time from camera images. However, this is a problem that I'm currently working on a solution for (see http://code.google.com/p/sentience/). > I know this is possible – I see it done again and again at AI conferences; Ah, well, appearances can be deceptive. There are many papers in computer vision in which you can see fancy 3D reconstructions produced from camera images. However, when you really get into the nitty gritty of how these work you'll usually find that they were either produced under highly contrived conditions or the result you can see is not statistically representative (i.e. you might get a good result, but only 20% of the time). Programs such as the CMU photopopup and photosynth appear impressive, but in the CMU case the reconstruction quality is poor (good enough for entertainment, but not much else) and the photosynth case they're still trying to reduce the huge amount of number crunching needed to produce the point cloud models (which can take hours or days with current computers). There is progress being made on 3D reconstruction using scanning laser rangefinders. This is the same kind of technology used in the DARPA chellenges, but it's not cheap and it's certainly not "off the shelf" in software terms. I think it will be possible to produce colour 3D models in real time from camera images using reasonably low cost off the shelf technology within about five years, but for the moment it remains as a kind of holy grail in computer vision. Once this happens then many new robotics applications will become possible. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=90535901-fc9d99
RE: [agi] Real-time Java-based vision libraries?
> Ah, well, appearances can be deceptive. There are many papers > in computer vision in which you can see fancy 3D reconstructions > produced from camera images. However, when you really get > into the nitty gritty of how these work you'll usually find that > they were either produced under highly contrived conditions or > the result you can see is not statistically representative (i.e. > you might get a good result, but only 20% of the time). Thanks Bob, For my purposes, I'm actually fairly comfortable with "highly contrived conditions", results that are very approximate or just robust object segmentation (rather than full/partial reconstruction). I share a lab with a Robocup team, and was toying with the idea of trying to adapt their C++ code and their highly contrived soccer field to my experiments. Unfortunately, though, Robocup vision would fail if you were to throw a yellow ball onto the field (instead of red): the systems aren't "open-ended" enough for my liking. > However, this is a problem that I'm currently working on a > solution for (see http://code.google.com/p/sentience/). This looks very interesting. How long (and on what sort of machine) does it take to process each stereo pair with your dense stereo correspondence algorithm? > These are all non-trivial problems and I don't know of any > libraries (java or otherwise) which "out of the box" perform > 3D reconstruction in real time from camera images. Thanks - it isn't looking too promising. It seems like I'm going to have to use bright lights, simple objects and then write some code of my own (or look into interfacing directly with a C++ system). -Ben - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=90806337-bf10cf
Re: [agi] Real-time Java-based vision libraries?
On 28/01/2008, Benjamin Johnston <[EMAIL PROTECTED]> wrote: > This looks very interesting. How long (and on what sort of machine) does it > take to process each stereo pair with your dense stereo correspondence > algorithm? The stereo correspondence takes about 50mS on any reasonably modern PC or laptop. Most of the processing time is in fact occupied by the SLAM algorithm which produces 3D grids. You can find a description of the stereo algorithm and a link to the code here http://code.google.com/p/sentience/wiki/StereoCorrespondence A possible alternative to stereo cameras which may become available within the next couple of years is something like the Z-cam ( http://www.youtube.com/watch?v=QfVWObYo-Vc ). It would be interesting to see this combined with Andrew Davison's monoSLAM. Anyway, by hook or by crook I think the kind of technology which you're looking for will arrive within the next few years, although for the present it remains just out of reach. Perhaps you should head for the nearest cryogenic chamber and re-emerge in five years time when the technology is ready for action. - Bob - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=90851447-6c72fa