Digital Signal Processing Reference
In-Depth Information
Multiple Objects Segmentation Using Graph Cut
Comparing with the single object segmentation, multiple objects segmentation as
a general case is investigated in our work. Based on the assumption that each
object is not overlapped, we convert multiple objects segmentation into several
sub-segmentation problems. For individual object, we construct a sub-graph for
the pixels belonging to its “Object Rectangle”, which is an enlarged rectangle of
bounding box to encompass the whole object and restricts the segmentation region.
Bi-label graph cut is employed to minimize the energy function and segments each
object. Experimental results using basic energy function with different and com-
bined features are shown in Fig.
5.4
a-c.
Modified Energy Function for Key View Segmentation
The segmentation quality using combined features in Fig.
5.4
c has outperformed
the ones using either single feature in Fig.
5.4
a,b. However, when the scenes contain
complex background, notable segmentation inaccuracy around the objects still exists
and leads to unsatisfactory results. These errors can be classified into two groups,
which are highlighted with rectangle and ellipse respectively, as shown in Fig.
5.4
c.
To tackle these two problems, we propose a modified energy function containing
two novelties:
background penalty with occlusion reasoning
is to handle the rect-
angle errors by refining data term in (
5.2
), and
foreground contrast enhancement
is to remove ellipse errors by refining smoothness term in (
5.3
). Based on the im-
portant observations that the focused object commonly appears in all the cameras,
and background regions around the object boundary are occluded by observing the
same scene from different perspective as shown in combined occlusion map
CO
t
Fig. 5.4
Segmentation results using basic energy function and refinement using modified energy
function: Basic energy function using (
a
) color; (
b
)depth;(
c
) combined color and depth; (
d
) results
using modified energy function;
top
: Reading sequence;
bottom
: Calling sequence
