Applications of Video Segmentation - Video Segmentation and Its Applications - page 147

Tabl e 6. 2 Application and related work
Applications Reference work

Visual surveillance and, traffic control [ 10 , 19 ]

Object based video coding and, event based scalable coding [ 16 , 22 ]

Three-D reconstruction(visual hull), video tooning and, rendering [ 5 , 20 ]

Augmented Reality, tourism, games and, surgery [ 17 , 24 , 25 , 27 ]

Content based Video Summarization [ 2 ]

Video conferencing and video phoning application [ 1 ]

scenarios, according to different application constraints and goals can be found in
[ 3 ]. There have been a lot of work in video segmentation using colour, motion based
methods [ 10 , 19 ].
We have summarized the list of potential applications and related work in
Tab le 6.2 .
In fact, there are other applications like video classification, which deals with
problem of categorizing a given videos sequence into one or predefined video genre.
For example one might be interested in finding all non identical duplicate videos
having some personality as a focus [ 21 ].
There is a paradigm shift from traditional segmentation to using depth [ 20 ],
attention and prior model information [ 14 ] in addition to color and motion-based
approaches [ 18 ].
Clearly, depth based approaches bear the potential discriminative power of as-
certaining whether the object is nearer of farer. We have proposed and evaluated a
GrabCut segmentation technique based on combination of colour and depth infor-
mation [ 20 ].
However, GraphCut techniques demand user initialization. As stated in [ 4 ], while
using GraphCut techniques, attention based models can be used instead of man-
ual initialization for segmentation process. The attention models can be based
on saliency map approaches, which leads to saliency-based segmentation model.
However, modeling visual attention models is still a challenging problem. Visual
attention models have been widely used in many applications.
In fact, to find what object is attended to and where the attention likely to be,
filtering and prioritizing the information is vital. This is analogous to the nature of
human fovea, which acts according to stimulus. There are two major computational
models of attention such as:

1. Bottom-up attention : It is based on combination of low level features which in-
clude both oriented as well as non-oriented features such as colour, contrast, and
orientation.
2. Top-down attention : It involves task dependent processing, which generally re-
quires some prior knowledge about the scene. In effect, the user attention is
guided by what he sees.

Itti et al. [ 8 ] has proposed a model for finding low-level surprise at every location
in video streams. The method correlates with gaze shifts of two human observers
watching complex video clips such as television programs.