Personal tools
  • Log in
You are here: Home Publications Simple features for low-level saliency
The GazeCom project is funded by the European Commission (contract no. IST-C-033816) within the Information Society Technologies (IST) priority of the 6th Framework Programme.
Document Actions

Simple features for low-level saliency

by Michael Dorr last modified 2008-06-12 13:13

Presented at the Third Scandinavian Workshop on Applied Eye-Tracking, Lund, Sweden

Eleonora Vig, Michael Dorr, Karl Gegenfurtner, and Erhardt Barth

We are interested in the low-level features contributing to saccade target selection in order to learn transformations designed to guide eye movements in an unobtrusive way. We assess how local spectral energy computed on a spatio- temporal multi-scale representation discriminates between attended and non-attended movie locations. If local spectral energy is predictive for attention then we can change the saliency distribution of a movie by changing energy. Other saliency measures may be more powerful but are not as easily changed. We recorded 40000 saccades while 54 subjects viewed 18 high-resolution videos of outdoor scenes. We grouped the saccades into short (0-4), medium (4-10) and long (10+ degrees) range. For obtaining non-fixated locations, we created a fixation map and randomly picked locations with a low gaze density on this map. Local spectral energy was computed on a range of spatio-temporal levels of a multiresolution pyramid both at fixated and non-fixated locations, obtaining thus for each location a feature vector containing the energy information on the different spatio-temporal levels. For our analysis, we used a support vector machine to learn (on these feature vectors) the difference in local spectral energy between the two classes. The kernel function was a Gaussian whose parameters were found by cross-validation.

We found that the average local spectral energy at salient locations was higher. Prediction improved by increasing the number of spatio-temporal levels. We also found that it varies with saccade amplitude: under all conditions, short range saccades were easier to predict (with an accuracy of 73%) than long ones (68%). Performing our analysis on feature vectors containing energy values computed on several spatio-temporal levels improves performance from 65% (obtained on only one level) to 73% (3 spatial / 4 temporal levels). We thus expect that gaze-contingent energy manipulation can have a gaze-guiding effect.


Powered by Plone CMS, the Open Source Content Management System

This site conforms to the following standards: