Saliency predicts change detection in pictures of natural scenes. Wright, M. J. Spat Vis, 18:413--430, 2005.
abstract   bibtex   
It has been proposed that the visual system encodes the salience of objects in the visual field in an explicit two-dimensional map that guides visual selective attention. Experiments were conducted to determine whether salience measurements applied to regions of pictures of outdoor scenes could predict the detection of changes in those regions. To obtain a quantitative measure of change detection, observers located changes in pairs of colour pictures presented across an interstimulus interval (ISI). Salience measurements were then obtained from different observers for image change regions using three independent methods, and all were positively correlated with change detection. Factor analysis extracted a single saliency factor that accounted for 62% of the variance contained in the four measures. Finally, estimates of the magnitude of the image change in each picture pair were obtained, using nine separate visual filters representing low-level vision features (luminance, colour, spatial frequency, orientation, edge density). None of the feature outputs was significantly associated with change detection or saliency. On the other hand it was shown that high-level (structural) properties of the changed region were related to saliency and to change detection: objects were more salient than shadows and more detectable when changed.
@article{ Wright05,
  author = {Wright, M. J. },
  title = {{{S}aliency predicts change detection in pictures of natural scenes}},
  journal = {Spat Vis},
  year = {2005},
  volume = {18},
  pages = {413--430},
  abstract = {It has been proposed that the visual system encodes the salience of
	objects in the visual field in an explicit two-dimensional map that
	guides visual selective attention. Experiments were conducted to
	determine whether salience measurements applied to regions of pictures
	of outdoor scenes could predict the detection of changes in those
	regions. To obtain a quantitative measure of change detection, observers
	located changes in pairs of colour pictures presented across an interstimulus
	interval (ISI). Salience measurements were then obtained from different
	observers for image change regions using three independent methods,
	and all were positively correlated with change detection. Factor
	analysis extracted a single saliency factor that accounted for 62%
	of the variance contained in the four measures. Finally, estimates
	of the magnitude of the image change in each picture pair were obtained,
	using nine separate visual filters representing low-level vision
	features (luminance, colour, spatial frequency, orientation, edge
	density). None of the feature outputs was significantly associated
	with change detection or saliency. On the other hand it was shown
	that high-level (structural) properties of the changed region were
	related to saliency and to change detection: objects were more salient
	than shadows and more detectable when changed.}
}

Downloads: 0