Colour in Context
Research group
Computer Vision Center

Computation of Saliency Maps using Psychophysical Measurements of Colour Induction

Applied Vision Association (AVA 2012) - 2012
Download the publication : VMO2012.pdf [67Ko]  
We built a low-level computational model of early processing with the aim of replicating brightness (Otazu -et al-,2008,VisRes,48,733-751) and colour (Otazu -et al-,2010,JoV,10:12,6) induction effects. The core of our model is a parametric 3-Dimensional weighting function operating on a wavelet decomposition. We call this function the Extended Contrast Sensitivity Function (ECSF), which was calibrated using psychophysically-measured chromatic induction data from an asymmetric colour and luminance matching experiment. Since the ECSF integrates the output of divisive normalisation mechanisms at different scales, we integrated it into a new computational saliency model, intended to predict eye fixation data on uncalibrated colour images. Surprisingly, this new ECSF-based saliency model achieved performance (measured in AROC and KL) similar to state-of-the-art computer vision algorithms computed on Judd-et al and Bruce & Tsotsos datasets (Murray-et al-,2011,CVPR,433- 440). In this work, we modified the ECSF parameters in order to optimize them with respect to the image datasets for saliency estimation. We used a genetic algorithm to search in the parameter space. Our results show that optimized parameters are virtually identical to those of the psychophysically-measured ECSF parameters. From a methodological point of view we conclude that, by adding psychophysical data that capture evolutionary advantages of the human visual system, one can improve the prediction performance of computational models of low-level visual tasks on large datasets. This conclusion is a counterpart to successful machine learning approaches that became ubiquitous in the computer vision community in the last decade, and an example of the usefulness of biologically inspired constraints.

BibTex references

@InProceedings\{VMO2012,
  author       = "Maria Vanrell and Naila Murray and Xavier Otazu and C. Alejandro Parraga",
  title        = "Computation of Saliency Maps using Psychophysical Measurements of Colour Induction",
  booktitle    = "Applied Vision Association (AVA 2012)",
  year         = "2012",
  abstract     = "We built a low-level computational model of early processing with the aim of replicating brightness
(Otazu -et al-,2008,VisRes,48,733-751) and colour (Otazu -et al-,2010,JoV,10:12,6) induction effects.
The core of our model is a parametric 3-Dimensional weighting function operating on a wavelet
decomposition. We call this function the Extended Contrast Sensitivity Function (ECSF), which was
calibrated using psychophysically-measured chromatic induction data from an asymmetric colour
and luminance matching experiment. Since the ECSF integrates the output of divisive normalisation
mechanisms at different scales, we integrated it into a new computational saliency model, intended
to predict eye fixation data on uncalibrated colour images. Surprisingly, this new ECSF-based saliency
model achieved performance (measured in AROC and KL) similar to state-of-the-art computer vision
algorithms computed on Judd-et al and Bruce \& Tsotsos datasets (Murray-et al-,2011,CVPR,433-
440). In this work, we modified the ECSF parameters in order to optimize them with respect to the
image datasets for saliency estimation. We used a genetic algorithm to search in the parameter
space. Our results show that optimized parameters are virtually identical to those of the
psychophysically-measured ECSF parameters. From a methodological point of view we conclude
that, by adding psychophysical data that capture evolutionary advantages of the human visual
system, one can improve the prediction performance of computational models of low-level visual
tasks on large datasets. This conclusion is a counterpart to successful machine learning approaches
that became ubiquitous in the computer vision community in the last decade, and an example of the
usefulness of biologically inspired constraints.",
  url          = "http://cic.uab.cat/Public/Publications/2012/VMO2012"
}

Other publications in the database

 © 2008 Colour in context Group | Computer Vision Center. All rights reserved | Contact webmaster |  Last updated: Monday 11 May 2009     eXTReMe Tracker