Visual Attention Modelling in a 3D Context - Novel 3D Media Technologies - page 99

frequency-tuned saliency and contrast-based concept [ 2 , 17 ]. Methods based on
Fast Fourier transformation offer a fast way to process information in real time. A
Spectral Residual approach (SR) [ 18 ] proposed by Huo et al. considers irregularity
as a clue from the smooth spectrum for saliency detection. Spectral Residual
approach (SR) depends on the observation that log spectra of different images
share alike information. Huo et al. assume that the average image has a smooth
spectrum and any object or area that causes an abnormality from the smooth
spectrum will grab the viewer

s attention. In SR approach image is first transformed
into frequency domain. Then the spectrum is smoothed and subtracted from the
original one. Finally the result is transformed back into the time domain. More
recently, Image Signature approach (IS) [ 19 ] based on Discrete Cosine transform
(DCT) and its variation [ 5 , 29 ] is also explored in detecting salient regions.

'

6.2.3 Detection Using Temporal Information

The main aim of video based saliency is to separate salient motions from the
background. In the context of video many approaches have been proposed. In
fact, many similar approaches from the still image domain extended to temporal
domain as well. For example, the Temporal Spectral Residual (TSR) [ 11 ] approach
is an extension of Spectral Residual (SR) [ 18 ] approach.
The Temporal Spectral Residual (TSR) [ 11 ] approach explores the possibility of
estimating salient regions by removing redundant information. Principally, the TSR
approach applies the SR approach on temporal slices (XT and YT planes). XT and
YT are the planes of image lines in a temporal domain. The final saliency map
comprises motion saliency, which is considered to represent saliency with the
motion information.
Another approach, the Temporal Image Signature (TIS) [ 29 ] is an extension of
Image Signature (IS) approach [ 19 ]. TIS approach explores the advantage of DCT
to estimate salient regions. In TIS approach instead of analyzing the Fourier
spectrum as in the Temporal Spectral Residual approach, DCT information for
saliency detection is used. TIS approach is discussed in more detail in the later
session.
Researchers have also suggested combining both spatial and temporal informa-
tion available in the scene in order to detect saliency regions. For example, Guo
et al. [ 9 ] proposed to use spatio-temporal information available in the image for
saliency detection. They introduce the idea of using the images phase spectrum of
Fourier Transform (PFT). They compared their results with those of Huo et al

s
algorithm [ 18 ] for various test images and calculated the minimum, maximum, and
average pixel difference. They concluded that the difference is marginal and can be
neglected.

'