iZoom: A Better Way to View Large Data Sets

1
iZoom A Better Way to View Large Data Sets?

• Mike Ashmore
• CPSC 462

2
Problem Eye Trackers Suck.

• Interactive applications reveal just how noisy
  and error-prone eye tracker data is
• 1 tracking accuracy on a good calibration
• Up to 30-40 pixels error on a high-resolution
  display
• High jitter of data just makes reliable selection
  that much harder

3
Problem Application Designers Suck.

• Teeny-tiny portions of the screen need to be
  accurately selected.
• Everything in WinAmp / XMMS 6-8 pixels
• Microsoft Word ruler bar (tabs, etc.) 2 pixels
• Image editing applications / WYSIWYG print layout
  tools 1 pixel
• But theyve been offering a magnifier tool for
  ages now! Perhaps theres a lesson to be learned
  

4
Solution 1 Make everything bigger so its
easier to select (the DUPLO approach)

• Since it works so well for image editors, well
  just magnify the whole screen.
• Big wins Faster, less fatiguing selection of UI
  elements (c.f. Fitts Law)
• Lossage Amount of information available
  decreases in proportion to the square of the
  magnification (2x magnification 1/4 as much
  information)

5
Solution 2 Let the display slide around on a
virtual desktop

• Big wins You get unlimited desktop real estate.
  Icons can be as chunky as you want for fast
  selection of targets within display.
• Lossage Very little context available for data.
  Again, 2x magnification of the virtual desktop
  leaves 3/4 of the desktop non-visible at all
  times.

6
iZoom Solution Gaze-Contingent Fisheye Displays

• Big win A portion of the screen is magnified,
  but context is still available in the periphery.
  Best of both worlds!
• Possible Problem Can people use it without
  getting motion sickness?
• Other Possible Problem Curiously, nobody else
  seems to have published much on this idea. Maybe
  it hasnt worked for anybody else, either.

7
(No Transcript)
8
The Experiment

• Simple selection task look at the window with
  the X
• Three conditions
• Non-fisheye (the control condition)
• Naïve fisheye
• Always-on version of lens
• Smart fisheye
• Lens only appears after detecting a certain
  intentness of fixation

9
(No Transcript)
10
Challenges

• Bad Calibration Frustrated test subject
• A Bug! Tobii never reports more than 200
  calibration data points. Are calibrations being
  truncated?
• Solution More sophisticated calibration routine
• Dont take calibration samples until were fairly
  certain subject is fixating in the right spot
• Redo calibration points that deviate too far from
  actual screen coordinates. Be fascist about it
  until every single point is near perfect

11
Challenges

• The enemy of fisheye performance the Gutwin
  effect (aka the oscillating lens of doom).
• Cause Non-intuitive mapping from control to
  display
• Look one inch to the left, expect to see the lens
  move exactly one inch to the left in data space
• Particularly troublesome because display and
  control are so confounded together

12
The Gutwin Effect

• Techniques to address this issue
• Remap control space to match distorted visual
  field
• Dont distort visual field at all (Miniotas /
  2004 expanding target zones)
• Grab and hold fisheye at position of initial
  fixation

13
Data Chart
14
Data Analysis

• Use of postgreSQL for outlier analysis
• Outlier criterion ? ? ??
• Future work (this evening, probably) Integrate
  SQL and R mathematical analysis package for
  instant build of results

15
Future Work

• More sophisticated data analysis.
• This is clearly a non-normal distribution. What
  is it, then? Gamma? Beta?
• Select-and-hold fisheye
• Center fixation at start of trials
• Counting task - when selection is not an issue,
  does fisheye improve performance (accuracy /
  speed) on detailed inspection tasks?