The system you are describing sounds very interesting. If it is a
'cheap' technique, able to work with consumer hardware etc, it may be
just the thing we're waiting for.
The Hawkins theory is indeed not tested on complex imagery but there
are different teams working on it. What is so special to the concept
(I think) is the fact that it combines invariant spatial AND temporal
pattern recognition and prediction. In every layer of the hierarchy,
it learns the most common spatiotemporal patterns and stores these at
each level. After some learning time, each level knows the most common
inputs. New input is matched to the stored patterns, classified, and
its 'score' is forwarded to the next layer using Bayesian network
calculation. You can see it as Hebbian learning, plus hierarchy, plus
Bayesian chance calculation to increase performance. The result is an
elegant solution that can be used for prediction of pretty complex
spatial and temporal patterns.
At a higher level in the hierarchy you could drop the hierarchical
propagation and let the nodes propagate to a larger subset to increase
complexity.
So what you get is a animal'ish short-term predicting agent. What the
system lacks, is long term prediction and a natural way of creating an
accessible 3D model of the environment. So the system your company is
working on would definitely be one welcome addition. But I'm just
undergraduate, new on the field, so please correct me on any errors
and tell me when to shut up ;)
Greetings, Durk
On 9/5/06, Bob Mottram <[EMAIL PROTECTED]> wrote:
On 05/09/06, Kingma, D.P. <[EMAIL PROTECTED]> wrote:
> A problem within the AI domain is that Vision has not been solved yet.
> The existing and functioning algorithms are mainly specialised into
> sub domains like face recognition etc. These are very nice but are not
> general enough to use in an arbitrary environment.
Fear not, for general purpose vision systems are on the way. I'm working on
general purpose vision at the moment, for a company in the US. The
materials transport robots produced by this company have the first general
purpose vision system, consisting of a few stereo cameras, and it works
remarkably well (reliably enough to navigate over long distances, within a
tollerance of about 3-5cm). Basically what the cameras see is reverse
engineered into a 3D model, not unlike the type of thing which you might
find in a computer game. Once this kind of information is available many
tasks become significantly easier than would be the case using other
sensors, such as lasers or ultrasonics. I expect this technology to get
cheaper and within a few years be within the realm of consumer applications.
I should say that these vision systems will work in *any* environment,
within the limitations of the dynamic range of the cameras. The only
situations where visual navigation doesn't work are those were the
environment is changing constantly, so that there's no repeatability in what
the robot sees, but otherwise they can cope with quite substantial
environmental changes (people or other vehicles wandering around).
I've not read Hawkins book, but in the lectures I've seen by him the kinds
of vision which he demonstrates are extremely simple "toy" problems using
binary images. It's also a very 2D approach, and yet we live in a
fundamentally 3D world. I would be very surprised if this kind of approach
worked with real camera images and all their inherent complexity and
uncertainty.
- Bob
________________________________
To unsubscribe, change your address, or temporarily deactivate your
subscription, please go to
http://v2.listbox.com/member/[EMAIL PROTECTED]
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
