We are planning to use ARToolkit for things similar.  We are just starting
on this project, and wondering how things worked.  One question
I have is: which system shall we use for our experiments:
PC, MAC-G4 or SGI?  How well the image processing algorithms
worked in the toolkit for all these systems?

Thanks
Semwal

> Dear All,
>
> I am looking for information on the ARMulti (sub) library and the multi
> sample program. Where can I find any documentation?
> Can multi be extended to use a larger number of markers? I had a look at
>  the file describing the markers distribution in space (marker.dat) and
> it seems so, but did anyone tried? Also, what does the matrix relative
> to each marker describe?
>
> I am thinking of using the ARToolkit to set up an computer vision based
> tracker with markers on the ceiling (similar to
> http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
> any other known similar system?
>
> Please let me know if you need more information.
>
> Regards,
> Enrico

Dear All,

I am looking for information on the ARMulti (sub) library and the multi 
sample program. Where can I find any documentation?
Can multi be extended to use a larger number of markers? I had a look at 
the file describing the markers distribution in space (marker.dat) and 
it seems so, but did anyone tried? Also, what does the matrix relative 
to each marker describe?

I am thinking of using the ARToolkit to set up an computer vision based 
tracker with markers on the ceiling (similar to  
http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there 
any other known similar system?

Please let me know if you need more information.

Regards,
Enrico

I hate to be negative, but unless you are planning on making some major 
changes to the AR Toolkit, this idea will not work.

We have tried to do it in the past, and it won't work with the existing 
ARTk for two reasons:
- accuracy.  The pose estimates returned from the ARTk are not accurate 
enough.  They are fine for doing what it was designed for (overlaying 
things on the fiducials), because it is not necessary to use real world 
coordinates for anything in that situation. But the poses are not 
actually very good in real-world coordinates, and the small jitters you 
often see in the objects orientation can become HUGE errors in the 
camera position if you invert the matrix to get the position of the 
camera relative to the markers.
- speed.  The template matching algorithm used basically works like 
this.  First, for each marker the ARTk knows about, create 4 templates 
(one at each orientation).  Second, when I get a marker, compare it to 
ALL known templates, and return the best match.  So, if you know about 
200 markers (say), then for each one seen you do a comparison to 800 
templates.  Now, if you see 10-20 markers in the scene (up on the 
ceiling) you must do 8000-1600 comparisons.  This causes the software 
to run VERY slowly.

We (and others, I think) have been trying to fix this, but it isn't 
ready for release yet.  We'll post something here when we have 
something, hopefully before the end of the year (we are also working on 
some automated autocalibration steps to go with it).

   blair

On Thursday, October 9, 2003, at 10:31  AM, Sudhanshu Semwal wrote:

>
> We are planning to use ARToolkit for things similar.  We are just 
> starting
> on this project, and wondering how things worked.  One question
> I have is: which system shall we use for our experiments:
> PC, MAC-G4 or SGI?  How well the image processing algorithms
> worked in the toolkit for all these systems?
>
> Thanks
> Semwal
>
>> Dear All,
>>
>> I am looking for information on the ARMulti (sub) library and the 
>> multi
>> sample program. Where can I find any documentation?
>> Can multi be extended to use a larger number of markers? I had a look 
>> at
>>  the file describing the markers distribution in space (marker.dat) 
>> and
>> it seems so, but did anyone tried? Also, what does the matrix relative
>> to each marker describe?
>>
>> I am thinking of using the ARToolkit to set up an computer vision 
>> based
>> tracker with markers on the ceiling (similar to
>> http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
>> any other known similar system?
>>
>> Please let me know if you need more information.
>>
>> Regards,
>> Enrico
>
>

Hi,

There have been a few papers written about building roof tracking systems
using ARToolkit. You must realise though that ARToolkit is really not
designed to perform this - it generates output values that appear to make
the target overlay the marker but these results are distorted if you view
the scene from any other angle.

The ART02 workshop had a couple of papers that you might be interested in. A
paper by Kalkusch et al at ART02 described a system using fiducial markers
on walls that were tracked to provide positioning and orientation.
(Kalkusch, M., Lidy, T., Knapp, M., Reitmayr, G., Kaufmann, H., and
Schmalstieg, D. Structured Visual Markers for Indoor Pathfinding. In 1st
Int'l Augmented Reality Toolkit Workshop, Darmstadt, Germany, Sep 2002 -
http://www.ims.tuwien.ac.at/media/documents/publications/SignPostFinal.pdf)

I also had a poster at ART02 which described problems with using ARToolkit
when integrated into scene graph software. It described some of the problems
with using the distorted camera calibration matrices that ARToolkit uses,
and ways of working around them (Piekarski, W. and Thomas, B. H. Using
ARToolKit for 3D Hand Position Tracking in Mobile Outdoor Environments. In
1st Int'l Augmented Reality Toolkit Workshop, Darmstadt, Germany, Sep
2002. - http://www.tinmith.net/papers/piekarski-art-2002.pdf)

At ISMAR03 just last week, we also presented a poster using ARToolkit
fiducial marker tracking. We used multiple shoulder looking cameras to
increase tracking probabilities and accuracy. However, ARToolkit will never
produce results as good as can be achieved using systems designed for the
task (such as the VisTracker by Foxlin and Naimark), but it does make
something relatively cheap and simple to deploy. (Piekarski, W., Avery, B.,
Thomas, B. H., and Malbezin, P. Hybrid Indoor and Outdoor Tracking for
Mobile 3D Mixed Reality. In 2nd Int'l Symposium on Mixed and Augmented
Reality, Tokyo, Japan, Oct 2003. -
http://www.tinmith.net/papers/piekarski-ismar-roof-2003.pdf)

Hope this is useful for you ...

regards,
Wayne

----------------------------------------------------------------------------
Wayne Piekarski - Researcher / Lecturer          pho:        +61-8-8302-3669
                                                 fax:        +61-8-8302-3381
Assistant Director - Wearable Computer Lab       mob:        +61-407-395-889
School of Computer and Information Science       ema:  wayne@c ..............
University of South Australia                    www: http://www.tinmith.net


----- Original Message -----
From: "Sudhanshu Semwal" <semwal@c ..........>
To: <enrico_groups@l ........>
Cc: <artoolkit@h ..................>
Sent: Friday, October 10, 2003 1:01 AM
Subject: Re: about using multiple markers for tracking


>
> We are planning to use ARToolkit for things similar.  We are just starting
> on this project, and wondering how things worked.  One question
> I have is: which system shall we use for our experiments:
> PC, MAC-G4 or SGI?  How well the image processing algorithms
> worked in the toolkit for all these systems?
>
> Thanks
> Semwal
>
> > Dear All,
> >
> > I am looking for information on the ARMulti (sub) library and the multi
> > sample program. Where can I find any documentation?
> > Can multi be extended to use a larger number of markers? I had a look at
> >  the file describing the markers distribution in space (marker.dat) and
> > it seems so, but did anyone tried? Also, what does the matrix relative
> > to each marker describe?
> >
> > I am thinking of using the ARToolkit to set up an computer vision based
> > tracker with markers on the ceiling (similar to
> > http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
> > any other known similar system?
> >
> > Please let me know if you need more information.
> >
> > Regards,
> > Enrico
>
>
>

Good point, Philip.

What I should have made more clear is that the pose estimates are 
inaccurate, but given the way the ARToolkit works, one usually does not 
notice (e.g., you might notice some graphics "wobble" a bit on their 
fiducial because of the poor orientation estimate).

However, when you invert the matrix, this error explodes:  you are now 
measuring the camera relative to the fiducial, and those small 
"wobbles" transform into huge position errors as the camera "swings" 
around the fiducial.

If you do it and look at the position and orientation of the camera, 
you will be quite amazed (we were).

  cheers,
   blair

On Tuesday, October 28, 2003, at 09:52  PM, Philip Lamb wrote:

> A bit of a late followup to the quoted discussion, but one point Blair 
> made needs to be clarified- there should be no loss of numerical 
> precision from inverting a coordinate matrix returned by ARToolKit, 
> since these coordinate matrices are actually homogenous coordinate 
> transforms.
>
> Homogenous coordinate transforms have some useful properties. Given a 
> HCT matrix T1:
> 	[ R11 R12 R13  p1 ]
> 	[ R21 R22 R23  p2 ]
> 	[ R31 R32 R33  p3 ]
> 	[  0   0   0   1  ]
> where (p1, p2, p3) is the position vector, and "R" is a 3x3 rotation 
> matrix (the matrix of unit vectors of the transformed coordinate 
> system referenced to the base coordinate system), then the inverse 
> T1^-1 is:
> 	[ R11 R21 R31  -p1 ]
> 	[ R12 R22 R32  -p2 ]
> 	[ R13 R23 R33  -p3 ]
> 	[  0   0   0    1  ]
>
> You can learn more about this from any basic robotics textbook. Note 
> that I am not implying that the ARToolKit is accurate enough for any 
> given application, but simply that there should be no loss of 
> numerical precision from inverting pose estimates that it returns.
>
> Phil.
>
>
> At 5:54 PM -0400 9/10/03, Blair MacIntyre wrote:
>> - accuracy.  The pose estimates returned from the ARTk are not 
>> accurate enough.  They are fine for doing what it was designed for 
>> (overlaying things on the fiducials), because it is not necessary to 
>> use real world coordinates for anything in that situation. But the 
>> poses are not actually very good in real-world coordinates, and the 
>> small jitters you often see in the objects orientation can become 
>> HUGE errors in the camera position if you invert the matrix to get 
>> the position of the camera relative to the markers.
>>
>>>> I am thinking of using the ARToolkit to set up an computer vision 
>>>> based
>>>> tracker with markers on the ceiling (similar to
>>>> http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is 
>>>> there
>>>> any other known similar system?
>>>>
>>>> Regards,
>>>> Enrico
>
>
> -- 
> --
> Philip Lamb
> B.E. (Hons.) M.Sc. (Hons.)
> PhD student,
> Human Interface Technology Laboratory New Zealand,
> University of Canterbury,
> Christchurch, NZ.

A bit of a late followup to the quoted discussion, but one point 
Blair made needs to be clarified- there should be no loss of 
numerical precision from inverting a coordinate matrix returned by 
ARToolKit, since these coordinate matrices are actually homogenous 
coordinate transforms.

Homogenous coordinate transforms have some useful properties. Given a 
HCT matrix T1:
	[ R11 R12 R13  p1 ]
	[ R21 R22 R23  p2 ]
	[ R31 R32 R33  p3 ]
	[  0   0   0   1  ]
where (p1, p2, p3) is the position vector, and "R" is a 3x3 rotation 
matrix (the matrix of unit vectors of the transformed coordinate 
system referenced to the base coordinate system), then the inverse 
T1^-1 is:
	[ R11 R21 R31  -p1 ]
	[ R12 R22 R32  -p2 ]
	[ R13 R23 R33  -p3 ]
	[  0   0   0    1  ]

You can learn more about this from any basic robotics textbook. Note 
that I am not implying that the ARToolKit is accurate enough for any 
given application, but simply that there should be no loss of 
numerical precision from inverting pose estimates that it returns.

Phil.


At 5:54 PM -0400 9/10/03, Blair MacIntyre wrote:
>- accuracy.  The pose estimates returned from the ARTk are not 
>accurate enough.  They are fine for doing what it was designed for 
>(overlaying things on the fiducials), because it is not necessary to 
>use real world coordinates for anything in that situation. But the 
>poses are not actually very good in real-world coordinates, and the 
>small jitters you often see in the objects orientation can become 
>HUGE errors in the camera position if you invert the matrix to get 
>the position of the camera relative to the markers.
>
>>>I am thinking of using the ARToolkit to set up an computer vision based
>>>tracker with markers on the ceiling (similar to
>>>http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
>>>any other known similar system?
>>>
>>>Regards,
>>>Enrico


-- 
--
Philip Lamb
B.E. (Hons.) M.Sc. (Hons.)
PhD student,
Human Interface Technology Laboratory New Zealand,
University of Canterbury,
Christchurch, NZ.

A bit of a late followup to the quoted discussion, but one point
Blair made needs to be clarified- there should be no loss of
numerical precision from inverting a coordinate matrix returned by
ARToolKit, since these coordinate matrices are actually homogenous
coordinate transforms.

Homogenous coordinate transforms have some useful properties. Given a
HCT matrix T1:
	[ R11 R12 R13  p1 ]
	[ R21 R22 R23  p2 ]
	[ R31 R32 R33  p3 ]
	[  0   0   0   1  ]
where (p1, p2, p3) is the position vector, and "R" is a 3x3 rotation
matrix (the matrix of unit vectors of the transformed coordinate
system referenced to the base coordinate system), then the inverse
T1^-1 is:
	[ R11 R21 R31  -p1 ]
	[ R12 R22 R32  -p2 ]
	[ R13 R23 R33  -p3 ]
	[  0   0   0    1  ]

You can learn more about this from any basic robotics textbook. Note
that I am not implying that the ARToolKit is accurate enough for any
given application, but simply that there should be no loss of
numerical precision from inverting pose estimates that it returns.

Phil.


At 5:54 PM -0400 9/10/03, Blair MacIntyre wrote:
>- accuracy.  The pose estimates returned from the ARTk are not
>accurate enough.  They are fine for doing what it was designed for
>(overlaying things on the fiducials), because it is not necessary to
>use real world coordinates for anything in that situation. But the
>poses are not actually very good in real-world coordinates, and the
>small jitters you often see in the objects orientation can become
>HUGE errors in the camera position if you invert the matrix to get
>the position of the camera relative to the markers.
>
>>>I am thinking of using the ARToolkit to set up an computer vision based
>>>tracker with markers on the ceiling (similar to
>>>http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
>>>any other known similar system?
>>>
>>>Regards,
>>>Enrico


-- 
--
Philip Lamb
B.E. (Hons.) M.Sc. (Hons.)
PhD student,
Human Interface Technology Laboratory New Zealand,
University of Canterbury,
Christchurch, NZ.

While there is no loss of numerical precision in the entire matrix, what
Blair was pointing out is that when you invert a matrix, errors in rotation
become errors in position, and vice versa. ARtoolkit traditionally tends to
have a lot of jitter in rotation (10-20 degrees typically) but has a more
stable position - by inverting the matrix you end up with the position
jumping around quite considerably. While theoretically there is no numerical
loss of precision, in practice the results are more difficult to use.

regards,
Wayne

----------------------------------------------------------------------------
Wayne Piekarski - Researcher / Lecturer          pho:        +61-8-8302-3669
                                                 fax:        +61-8-8302-3381
Assistant Director - Wearable Computer Lab       mob:        +61-407-395-889
School of Computer and Information Science       ema:  wayne@c ..............
University of South Australia                    www: http://www.tinmith.net


----- Original Message -----
From: "Mark Billinghurst" <grof@h ..................>
To: <artoolkit@h ..................>
Sent: Friday, October 31, 2003 6:14 PM
Subject: Re: about using multiple markers for tracking


>
> A bit of a late followup to the quoted discussion, but one point
> Blair made needs to be clarified- there should be no loss of
> numerical precision from inverting a coordinate matrix returned by
> ARToolKit, since these coordinate matrices are actually homogenous
> coordinate transforms.
>
> Homogenous coordinate transforms have some useful properties. Given a
> HCT matrix T1:
> [ R11 R12 R13  p1 ]
> [ R21 R22 R23  p2 ]
> [ R31 R32 R33  p3 ]
> [  0   0   0   1  ]
> where (p1, p2, p3) is the position vector, and "R" is a 3x3 rotation
> matrix (the matrix of unit vectors of the transformed coordinate
> system referenced to the base coordinate system), then the inverse
> T1^-1 is:
> [ R11 R21 R31  -p1 ]
> [ R12 R22 R32  -p2 ]
> [ R13 R23 R33  -p3 ]
> [  0   0   0    1  ]
>
> You can learn more about this from any basic robotics textbook. Note
> that I am not implying that the ARToolKit is accurate enough for any
> given application, but simply that there should be no loss of
> numerical precision from inverting pose estimates that it returns.
>
> Phil.
>
>
> At 5:54 PM -0400 9/10/03, Blair MacIntyre wrote:
> >- accuracy.  The pose estimates returned from the ARTk are not
> >accurate enough.  They are fine for doing what it was designed for
> >(overlaying things on the fiducials), because it is not necessary to
> >use real world coordinates for anything in that situation. But the
> >poses are not actually very good in real-world coordinates, and the
> >small jitters you often see in the objects orientation can become
> >HUGE errors in the camera position if you invert the matrix to get
> >the position of the camera relative to the markers.
> >
> >>>I am thinking of using the ARToolkit to set up an computer vision based
> >>>tracker with markers on the ceiling (similar to
> >>>http://www.bbc.co.uk/rd/pubs/papers/paper_05/paper_05.html). Is there
> >>>any other known similar system?
> >>>
> >>>Regards,
> >>>Enrico
>
>
> --
> --
> Philip Lamb
> B.E. (Hons.) M.Sc. (Hons.)
> PhD student,
> Human Interface Technology Laboratory New Zealand,
> University of Canterbury,
> Christchurch, NZ.
>