Self proposed projects for GSOC 2010

[Vishal]

Hello all,
I want to take part in GSOC 2010.
Last year there was an option for students to propose their own
project and I wanted to ask what are the relevant fields in which the
projects can be proposed by the students who are willing to apply for
OpenCog.
What are the asic fields in which the projects will be entertained?
I will be doing a project in Artificial Intelligence myself which will
be completed before the coding of GSOC starts.
So maybe I could do aproject and then build upon itin GSOC.
Cheers,
Vishal.

[Ben Goertzel]

We are open to anything, but projects that make use of the OpenCog or
OpenBiomind codebases are preferred ... and ideal are projects that
leave something useful for others to build on and apply after the
summer is done...

ben g

> --
> You received this message because you are subscribed to the Google Groups "OpenCog Summer of Code Announcements & Discussion List" group.
> To post to this group, send email to openc...@googlegroups.com.
> To unsubscribe from this group, send email to opencog-soc...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/opencog-soc?hl=en.
>
>

--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
Director of Applied Research, Singularity Institute for AI
Chairman, Humanity+
External Research Professor, Xiamen University, China
b...@goertzel.org

"Persistence and determination are omnipotent." -- Calvin Coolidge

[Joel Pitt]

Hi Vishal,

You can see some of the projects under the GSoC wiki category:

http://www.opencog.org/wiki/Category:GSoC

There are also a lot of ideas available here:

http://www.opencog.org/wiki/Ideas

Joel Pitt, PhD
http://ferrouswheel.me | http://opencog.org
+64 21 101 7308

[Vishal]

Thanks for the replies.
I am not much familiar with the opencog framework.
But I have taken a tour of it on the opencog webpage.
I was thinking of a project which to link the actuators with the
opencog framework.
I was thinking of implementing computer vision using openCV as a
module.
I think this is a project on which others can surely build upon.
Vishal.

[Vishal]

I will be working on openCV this semester also there will be some work
on the Nao robot too.
I am sure we will be doing something for its movement and make a face
recognition software for it.
So I may will be able to use the Nao owned by my college for the
experiments provided it is avaiilable in the summers.
Vishal

[Ben Goertzel]

Yes, that would be quite interesting to us, since we have some
students in China working on connecting OpenCog to a Nao robot as
well...

Face recognition is worthwhile, but I'd be more interested in stereo
vision from the Nao's eyes (though maybe that is too hard for a GSoC
project!!)

ben

> --
> You received this message because you are subscribed to the Google Groups "OpenCog Summer of Code Announcements & Discussion List" group.
> To post to this group, send email to openc...@googlegroups.com.
> To unsubscribe from this group, send email to opencog-soc...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/opencog-soc?hl=en.
>
>

--

[Nil Geisweiller]

On Wed, Feb 3, 2010 at 3:34 PM, Ben Goertzel <b...@goertzel.org> wrote:
> Yes, that would be quite interesting to us, since we have some
> students in China working on connecting OpenCog to a Nao robot as
> well...
>
> Face recognition is worthwhile, but I'd be more interested in stereo
> vision from the Nao's eyes (though maybe that is too hard for a GSoC
> project!!)

maybe it is not that hard, or at least it seems people are trying

http://www.starlino.com/opencv_qt_stereovision.html
http://en.wikipedia.org/wiki/OpenCV#Applications

Nil

[Ben Goertzel]

It's not hard "in principle" but that sort of work is specific to the
cameras involved...

The Nao's cameras are not that great, and they're arranged one above
the other, rather than horizontally, which is odd...

Bumblebee provides a great stereo vision system, with software tuned
to their particular hardware, but it's not usable with the Nao's
cameras ---- and also is not usable with the Nao, because the
Bumblebee cameras are too heavy for the Nao to carry, and Bumblebee
uses USB for power and data, and Nao lacks a USB port...

ben

"Before something is a breakthrough, it's a
crazy idea" -- Peter Diamandis

[Bob Mottram]

On 3 February 2010 15:41, Ben Goertzel <b...@goertzel.org> wrote:
> The Nao's cameras are not that great, and they're arranged one above
> the other, rather than horizontally, which is odd...
>
> Bumblebee provides a great stereo vision system, with software tuned
> to their particular hardware, but it's not usable with the Nao's
> cameras ---- and also is not usable with the Nao, because the
> Bumblebee cameras are too heavy for the Nao to carry, and Bumblebee
> uses USB for power and data, and Nao lacks a USB port...

I have some stereo vision code here, which could be adapted for the Nao.

http://code.google.com/p/libv4l2cam/

This is a sparse feature based method, but it's fast and the results
look like this.

http://www.youtube.com/watch?v=EUcLAarcj7U

There are plenty of other stereo algorithms to try, but very few that
actually work without making a lot of implausible assumptions. The
Birchfield algorithm used within OpenCV looks good in papers, but from
my tests in practice is unusable (more noise than signal).

An interesting recent approach is RelativeSLAM
http://www.robots.ox.ac.uk/RelativeSLAM/

[Ben Goertzel]

Interesting, Bob ... if no GSoC student wants to adapt your code to
the Nao, maybe we can get a Chinese student to do it for a MS project
next year...

ben

[Bob Mottram]

On 3 February 2010 17:28, Ben Goertzel <b...@goertzel.org> wrote:
> Interesting, Bob ... if no GSoC student wants to adapt your code to
> the Nao, maybe we can get a Chinese student to do it for a MS project
> next year...

The V4L2stereo code is very simple stuff which was primarily written
to run on a low powered DSP, so it could probably be easily adapted to
many platforms. Although the ranges are always subject to a certain
amount of noise it's possible you could use the information as a prior
when generating hypotheses about the shape of the surrounding space.
Getting good visual perception is probably a matter of doing several
different sorts of processing in parallel, of which stereoscopy is
only one, then integrating the results at a higher level.

[Ben Goertzel]

But the basic functionality of the code is to take input from two
cameras and fuse them into a single image including depth information?

[Bob Mottram]

On 3 February 2010 17:43, Ben Goertzel <b...@goertzel.org> wrote:
> But the basic functionality of the code is to take input from two
> cameras and fuse them into a single image including depth information?

Yes. In this case the depth information is sparse, as matched edge
features, although dense methods also exist.

[Ben Goertzel]

What's the advantage of sparse depth information?

Faster to compute?

[Bob Mottram]

Incidentally, if any students are thinking of tackling stereo vision
as part of GSoC I'd recommend the following book, which gives a good
overview of the issues and algorithms.

An Introduction to 3D Computer Vision Techniques and Algorithms
Boguslaw Cyganek & J. Paul Siebert
ISBN 9780470017043
http://www.amazon.com/Introduction-Computer-Vision-Techniques-Algorithms/dp/047001704X/ref=sr_1_1?ie=UTF8&s=books&qid=1265220181&sr=8-1

[Bob Mottram]

On 3 February 2010 17:59, Ben Goertzel <b...@goertzel.org> wrote:
> What's the advantage of sparse depth information?
>
> Faster to compute?

The main advantage is speed. Dense methods take longer to compute,
and in practice are often implemented in hardware as is the case for
commercial stereo cameras such as Point Grey and Videre.

[Murilo Saraiva de Queiroz]

I wonder how large scale availability of very cheap time-of-flight 3D cameras from Microsoft's Project Natal will render stereo vision almost irrelevant. 

The videos showing raw 3D data provided by Project Natal camera are very impressive: basically you receive a 2-centimeters resolution voxel field, updated 30 times per second (see  http://www.youtube.com/watch?v=-_UzcnTYqc4 - specifically the section beginning at 2:00). And by definition the problem of assigning a depth for each pixel in the image simply disappears... 

All the processing is done by custom hardware, and the final product will cost less than US$ 100. Microsoft promises standard USB connectivity, including drivers and a SDK for the PC.

It's funny that almost everybody (including the "specialized" press and magazines like New Scientist) is thinking about Project Natal as a "Wii-killer", when cheap TOF cameras are much more important than any human-interface application. I don't care about dancing in front of my TV while playing videogames, all I want is a cheap 3D vision system that simply works. :-) 

muriloq

--

You received this message because you are subscribed to the Google Groups "OpenCog Summer of Code Announcements & Discussion List" group.
To post to this group, send email to openc...@googlegroups.com.
To unsubscribe from this group, send email to opencog-soc...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/opencog-soc?hl=en.

--
Murilo Saraiva de Queiroz, MSc.
Senior Software Engineer  
http://www.vettalabs.com
http://www.tecnologiainteligente.com.br
http://www.acalantoemcasa.com.br

[Ben Goertzel]

hard to say until we get the camera and see how good it is...

At very least, it should provide a way to guide stereo vision systems,
making their job way easier... at best it will obsolete them, right?

--

[Bob Mottram]

On 5 February 2010 18:30, Murilo Saraiva de Queiroz <mur...@gmail.com> wrote:
> I wonder how large scale availability of very cheap time-of-flight 3D
> cameras from Microsoft's Project Natal will render stereo vision almost
> irrelevant.

You may well be right about this. In fact, I'm expecting that this
will be the case a few years from now, and that much of the
conventional stereo vision will be rendered obsolete.

However, demos involving 3D computer vision shouldn't necessarily be
taken too literally. It's possible to pull all sorts of tricks in a
demo to make performance look better than it actually is, so I'll wait
until I can actually try one of these devices myself before I come to
a firmer conclusion about how relevant it might be for robotics.

[Murilo Saraiva de Queiroz]

Regarding the demos, it seems their main problem now is correct, real-time detection of player's body posture and movement. It seems they are almost there (it's a massive application of machine learning). The requirements for navigation and other simpler robotic tasks aren't as sophisticated. Even if the resolution isn't 1-2 cm voxels but something much worse (e.g. 10 cm), and the range is kept in 1-2 m tasks like navigation inside an unknown indoor environment would become much easier.

Anyway I agree, until we put our hand in one of these it's just speculation... But it's pretty exciting stuff! ;-)

--

You received this message because you are subscribed to the Google Groups "OpenCog Summer of Code Announcements & Discussion List" group.
To post to this group, send email to openc...@googlegroups.com.
To unsubscribe from this group, send email to opencog-soc...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/opencog-soc?hl=en.

[Ben Goertzel]

Yes, it sounds extremely promising ... and increases my belief that we
should focus our efforts on cognition as sensation/action are being
more effectively solved by other people...

ben

On Sat, Feb 6, 2010 at 7:15 AM, Murilo Saraiva de Queiroz

--