A few weeks ago I asked what kind of constant error you all see in
your eye tracking data.
To summarize the responses, three people responded describing the
type of errors that they saw. All three respondents seem to observe
some kind of constant error in their data. Tony Renshaw finds that
the constant error is distributed more vertically than horizontally,
and co-authored a paper relevant to the topic. Hari Narayanan finds
that the constant error varies across subjects and regions of the
screen, but it seems as if he finds it to be consistent within a
participant (these findings are all consistent with mine). He
adjusts for the error with some post-calibration auto-correction.
Stefanie Kraft found similar error, but it was partly because the
brightness of the calibration background deviated from that of the
stimuli. She also corrected for constant errors using drift
corrections.
Keith Karn also responded and pointed out a number of potential
technical issues that may contribute to the constant error I am
observing.
Thank you all for your replies.
Anthony Hornof
-----Original Message-----
From: Anthony Hornof [mailto:[log in to unmask]]
Sent: Thursday, May 09, 2002 1:23 AM
To: [log in to unmask]
Subject: [EM_LIST] What kind of constant errors do you see?
I am using a particular eyetracker and see some pretty regular
patterns of constant error within each subject, but that vary across
subjects. By constant error, I mean that the eye tracker
consistently records fixations at location L2 when I am reasonably
confident that the gaze was really at L1. I am curious if any of you
can concisely describe the sorts of constant errors you see within or
across subjects, and within or across different regions of the
screen. If I get enough interesting replies, I'll compile them and
re-post.
1. What kind of constant error do you see?
2. What eye tracker do you use?
3. Do you have any other comments on this topic?
Best wishes,
Anthony Hornof
University of Oregon
http://www.cs.uoregon.edu/~hornof/
--------------------------------------------------------------------------
From: "Renshaw, Tony [IES]" <[log in to unmask]>
Date: Thu, 9 May 2002 11:05:05 +0100
Hi,
I am a Phd student at Leeds Metropolitan University (England).
I recently conducted an experiment using an ASL 504 pan/tilt eye
tracker in which I asked eight subkjects to gaze at nine objects
evenly distributed on a vdu screen for several seconds (>10 secs). I
noticed that the mean location of the eye position co-ordinates as
reported by the equipment was to the left and below the known
position of the objects being looked at. If this is the kind of thing
you are looking at I have more information available.
Yours
Tony Renshaw
[log in to unmask]
[Tony also forwarded a paper he co-authored entitled "The impact of
object dimension on eye gaze." The paper demonstrates that recorded
fixations tend to be distributed more greatly in the vertical than in
the horizontal, but not affected by an object's size.]
--------------------------------------------------------------------------
Date: Thu, 09 May 2002 09:26:24 -0400
From: "Karn, Keith S" <[log in to unmask]>
Anthony -
This sounds like a calibration error. What sort of calibration are you
doing? Another possibility is that there is a good calibration initially,
then there is some change in the setup (e.g., a head-mounted tracker
slipping on the head).
Just some simple thoughts that you have probably already considered.
Keith Karn
----------
Hi Keith,
Thanks very much for your reply. I'm using the LC Technologies
Eyegaze and I don't think it's a calibration error because it occurs
immediately after the calibration (I use the 9-point calibration
shown at http://eyegaze.com/doc/eds.htm) and because the eye tracker
specs allow for some constant error even with accurate calibration.
There is *some* deterioration over time, but not with all
participants, and even then, the constant error persists immediately
after re-calibration. But the issues you raise are exactly those
that I'm trying to examine. I'd like to figure out how much of the
error is a calibration error, and how much is constant error that
will persist after a correct calibration, for which the challenge
becomes doing some post-hoc analysis to remove that error.
Best wishes,
Anthony
----------
Anthony -
By definition, shouldn't there be zero error after calibration? This should
at least be true on the spots on the screen on which you fixate during the
calibration procedure (I assume there is some interpolation between these
points). Could the errors be the result of a parallax problem where you are
calibrating at one distance from the observer then noticing an error when
fixating at another distance?
Keith K.
--------------------------------------------------------------------------
From: "N. Hari Narayanan" <[log in to unmask]>
Date: Thu, 9 May 2002 08:56:43 -0500 (CDT)
Anthony
I use an SMI EyeLink and we had this problem, which we have improved by
adding our own calibration step after the system calibration and
validation steps, in which selected points from the stimulus are displayed
and accuracy of calibration is verified manually by the experimenter
(and the whole procedure repeated as necessary) before every trial.
This may not make sense if yours isn't SMI EL.
I'd be interested in seeing other responses - so do compile and post.
Best,
Hari
----------
Hi Hari,
...
Your manual post-calibration calibration process sounds very
interesting. I think we are trying to develop a means of doing
roughly the same thing but somewhat automatically, and in which the
corrections are made in the post hoc analysis of the data.
Can you tell me a little more about the type of error that you saw?
For example, was everything shifted to the left, or usually down, or
did it vary for different regions of the screen, and did it vary from
participant to participant?
...
Best wishes,
Anthony
----------
From: "N. Hari Narayanan" <[log in to unmask]>
Date: Fri, 10 May 2002 03:53:08 -0500 (CDT)
>Can you tell me a little more about the type of error that you saw?
>For example, was everything shifted to the left, or usually down, or
>did it vary for different regions of the screen, and did it vary from
>participant to participant?
There was no predictable pattern. Sub-to-sub variation, shifts could
be up, down left or right, there were also variations across regions.
We thought about doing some auto-correction but given the variations decided
that reducing and in many cases avoiding this problem by add'l calib.
was better.
One could however think about dividing the screen into regions and
have the subject look at points in each both prior and after a
stimulus and do some auto correction based on an estimate of drift
using data from these two calib steps.
See you soon,
Hari
--------------------------------------------------------------------------
Date: Tue, 14 May 2002 09:27:39 +0200
From: stefanie Kraft <[log in to unmask]>
Hi,
we had a similar problem when we started recording binocular eye data
with the SMI eyetarcker. Here, the main problem was the calibration.
The brightness of the calibration background deviated form the
backgound of the stimuli. This lead to changes in the size of the
pupil. The eyes seemed to diverge over the time.
We still observe small deviations within a measurement but now these
result from movements of the head set.
Regards
Stefanie Kraft
----------
Date: Tue, 14 May 2002 17:19:23 +0200
From: stefanie Kraft <[log in to unmask]>
Please note that I should have mentioned that the eye tracker we use is
the eye link system. The problem we had was user mistake. The
deviations we see now are mostly within the tolerance spectrum and
can be minimized by using as many drift corrections as possible.
Stefanie Kraft
--
EYE-MOVEMENT mailing list ([log in to unmask])
N.B. Replies are sent to the list, not the sender
To unsubscribe, etc. see http://www.jiscmail.ac.uk/files/eye-movement/introduction.html
Other queries to list owner at [log in to unmask]
|