A robust algorithm for reading detection

1
A robust algorithm for reading detection

• ACM International Conference Proceeding Series
  Vol. 15
• Proceedings of the 2001 workshop on Perceptive
  user interfaces
• Christopher S. Campbell , Paul P. Maglio  
• IBM Almaden Research Center, San Jose, CA

2
outline

• Introduction
• Utility of detecting reading
• System and implementation
• Experiment and result
• Discussion
• Future work
• My thought

3
INTRODUCTION

• Developing a method for recognizing when users
  are reading text based on eye movement data.
• Compared to a simple detection algorithm, our
  algorithm reliably, quickly, and accurately
  recognizes and tracks reading.

4

• eye movement data analysis can be classified into
  three different levels
• highly detailed low-level micro-events
• low-level intentional events, and
• coarse-level goal-based events.
• ultimate aim is to track text the user reads to
  infer user interests and goals.

5
Utility of detecting reading

• In many such gaze-based systems, interest in some
  display object is determined by a fixation
  threshold.
• Reading detection provides a much more precise
  means for determining user interest because it
  can determine the level of user interest(
  reading, skimming, scanning).

6
System and implementation
Example pattern of eye movements for Participant
1 reading a paragraph of text
7

• Reading behaviors depend on several factors,
  including text difficulty, word length, word
  frequency, font size and color, user distance to
  display, and individual differences.
• For example, the text becomes more difficult to
  comprehend, fixation duration increases, and the
  number of regressions increase.

8

• reading detection system relies on three
  mechanisms
• coarse or quantized represented of eye-movements,
• Pooled evidence based detection,
• mode switching

9
quantized represented of eye-movements

• the eye-movements in both x and y positions are
  quantized (averaged) over 100 ms intervals.
• This process removes some of the inaccuracy of
  current eye-tracking hardware and reduces the
  influence of micro-saccades.

10
Pooled evidence based detection

• evidence of reading is accumulated until it
  crosses a threshold value.
• This is done by the right and de-incrementing
  when the eye moves to the left.

11
mode switching

• the evidence reaches a threshold, then reading
  is detected and the mode is switched to reading
  from scanning.
• ?Mode switching allows us to essentially
  interpret the same eye movements differently
  based on changes in context. For example
  Revisit.

12
(No Transcript)
13

• 10 5 -5 10 10 30 (Reading mode)
• A single scan jump event will send system back
  into scanning mode.

read forward
skim forward
skim jump
skim forward
skim forward
Scan jump
14

• Another algorithm - Jacob's fixation recognition
  algorithm, but the goal of Jacobs method differs
  from our goal of recognizing reading.
• To compare Jacob's method with our own, we
  suppose a series of say three fixations to the
  right signal that reading is detected.

15
EXPERIMENT

• we compare our algorithm with the extended
  Jacobs algorithm to determine which performs
  faster and more reliable.
• experiments investigated recognition accuracy
  (Experiment 1) and recognition speed (Experiment
  2).

16
EXPERIMENT 1

• Participants were instructed to perform two
  tasks
• (a) quickly search for a target icon on a screen
  full of distracter icons, and
• (b) carefully read a text passage and then answer
  a multiple-choice
• These two tasks were presented in random order,
  and eye movements were recorded.

17
(No Transcript)
18
EXPERIMENT 2

• Participants were presented with a series of
  passages (each containing about four sentences)
  that varied in reading difficulty.
• used a single factor (text difficulty)
  within-subjects design with two conditions (easy
  or difficult) and fifteen trials per condition
  (30 trials total).

19
(No Transcript)
20
Discussion

• results show that the pooled evidence algorithm
• faster at detecting reading
• high (nearly 100) accuracy rate
• Reliable across different participants and styles
  of text.

21

• The results also show that text difficulty did
  not influence gaze patterns as much as expected
  -- only the number of regressions varied with
  difficulty.

22
Future work

• Detecting skimming in reading mode, if the
  distance is greater than threshold then the words
  are classified as skimmed.
• Personalization We will include parameters that
  adapt to individual reading speeds and abilities
  by adjusting parameters

23

• Context information to constrain reading
  detection
• location of text on the screen
• the size of the font
• the content of the text on the screen
• Other context information

24
My thought

• ??????????????,?????????????????,????????,????????
  ????????,???????
• ?machine learning??????????????????
• ???future work??????????????????,?????????????????
  ??????
• ???????????(??????????????),??????????????????????
  ?