Digital Signal Processing Reference
In-Depth Information
As an important low-level feature, motion can provide the otherwise missing
semantic information in cases where uniform motion is expected. In order to ex-
tract the moving objects, motion estimation methods are needed especially when
change detection masks have been shown to be ineffective. An interesting work on
moving object segmentation can be referred to [
2
]. The core of this algorithm is
an object tracker that matches a two-dimensional (2D) binary model of the object
against subsequent frames using the Hausdorff distance. To achieve this goal, the
first step is to detect a dominant global motion that can be assigned to the back-
ground based on the six-parameter affine transformation. An object tracker based
on Hausdorff distance is then established to measure the temporal correspondence
of objects and enhance the robustness to noise and changes in shape in the video
sequence.
1.3
Technological Trends for Image/Video Segmentation
Most past research activities on video segmentation have relied on two principles of
spatial (i.e., image) and temporal segmentation. If we treat the motion cue as one of
the low level features such as intensity, color, and texture, many image segmentation
algorithms can be easily extended to video segmentation. For example, to segment
a moving object out from a video clip, a 3D graph cut was presented to partition
watershed presegmentation regions into foreground and background while preserv-
ing temporal coherence. For each frame, the segmentation in each tracked window
is refined using a 2D graph cut based on a local color model [
36
]. In this section,
we will address the following trends for segmentation algorithm especially for the
spatial domain segmentation.
1.3.1
Towards 'Good' Segmentation
An emerging trend is to answer the question “What is a good partition for an
image?” An interesting work in the current literature is to group pixels into “su-
perpixels”, which are local, coherent, and which preserve most of the structure
necessary for segmentation at the scale of interest [
13
,
37
]. To generate the super-
pixel map, the Ncut segmentation algorithm is used by incorporating the contour
and texture cues. To find the “good” segmentation, the gestalt grouping cues, such
as contour, texture, brightness, and good continuation are combined in a principled
way. A linear classifier is trained to combine these features.
An example of superpixel segmentation is shown in Fig.
1.5
, which has the num-
ber of superpixels 200. The original image flower is shown in Fig.
1.5
a, which has
the superpixel map given in Fig.
1.5
b. A result of segmentation can be found in
Fig.
1.5
c, which shows that distinct improvement can be achieved with respect to
those classic methods.
