0

I would like to compare two eye-tracking scanpaths. The eye-tracking results in a sequence of labels that the observer look at, for a division of an image into labeled tiles (rectangular regions). We also know from the eye-tracking at what time, and for how long the eye look at tile N.

The Levenshtein or string edit distance works fine, as long as the timing of the fixations are not taken into account. For example, f user 1 looks at the tiles "AKPLA", and user 2 looks at tiles "ATPLB" the string edit distance will be 2, but user 2 might look at "P" in a much longer time than user 2.

Any ideas of how to improve the distance measure to measure timing differences as well? (note that the algorithm is not restricted to character strings, it works equally well with arrays of integers).

melhosseiny
  • 9,992
  • 6
  • 31
  • 48
Stefan
  • 622
  • 1
  • 5
  • 17
  • This might be better suited to http://cogsci.stackexchange.com/ - if you think that makes sense don't double post, flag this post and ask to have it moved. – Flexo May 10 '12 at 09:46
  • Well, I just tried that now -- but I cannot use any of my tags, so I'm not sure that is the right forum. Thanks anyway. – Stefan May 11 '12 at 08:34
  • What is the status of this question as of now? – melhosseiny Apr 28 '13 at 00:08

1 Answers1

0

An eye-tracking scanpaths originally would be a time-series. Transforming your time-series into a string only containing the labels where the person looked at leads to loosing information about time.

Thus if you want to take the time into account, you have to either work with the original time-series or take the time into account for your transformation.

For example: You could for every ten seconds give the laben where the person looked at on average. It could be "AAAAKPLAA" as compared to "AATTTPLBB". In this case you could use the Edit Distance and it would take into account how long someone looked where.

You could also simply work on the original time-series of eye-tracking which - as I assume - contains a time-stamp and a position. Then you could use Dynamic Time Warping to estimate the dissimilarity.

Anyhow, this is a very broad question and probably of no relevance for you any longer. If you could post the answer that you found yourself, it would be great.

Nikolas Rieble
  • 2,416
  • 20
  • 43
  • 1
    Thank you for your, although rather late, answer. I decided not to use the timing information, as a time series analysis was outside the scope of the project. I think your suggestion of coding the timing into the strings requires that you somehow calibrate the influence of the timing information relative to the influence of shift from one tile to another. I guess that it will be possible to do so, but it will require some experiments. – Stefan Jul 26 '16 at 13:51