Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 26

A

A C B

?

B

Layer boundary
b.

Possible layer boundaries
a.

Fig. 1 Illustration of moving regions ( A , B ) and occluded region ( C ). (a) The probable layer
boundaries are determined by extending the moving region to the occluded region; (b)
There is no occlusion region between layers A and B .
cannot find out the layer order according to the distinct edges of the motion layers.
The layer boundaries are not the same as the occluding edges. The layer bounda-
ries involve the occluding edges, but the layer boundaries are not always the oc-
cluding edges. It is infeasible to infer the layer order only by the layer boundaries.
We can therefore conclude that the occlusion relationship hides behind the oc-
cluded regions, and identifying the occluding edges can reveal the occlusion rela-
tionship. The optical flow computation can usually identify the coarse occlusion
regions as a by-product [15], which will be adopted in this paper.
The subspace segmentation approach described in section 2 is carried out on a
given optical flow field instead of the image intensities. Due to the errors from the
optical field (e.g. aperture problem etc.), each resulting motion layer contains two
kinds of artifacts: (1) small isolated regions with texture and (2) dark holes over the
image plane. It can be observed that a single hole in the middle of a foreground
layer runs through to the background layer. Similar problems also exist in the oc-
cluded regions. Moreover, the resulting boundaries of motion layers and their
neighborhood are, in general, highly unreliable areas. Therefore, the segmentation
by the subspace segmentation method and the occluded regions detected by the
optical flow computation cannot offer a valid solution to the above two problems.
Consider the neighborhood of the layer boundaries. It can be observed that the
occluded regions are involved in the neighborhood of the layer boundaries as
shown in Fig.1a. The edges' neighborhood contains the wealthy intensity struc-
tures of image. This can provide us sufficient cues to find the layer boundaries and
occluding edges. We rephrase the problem of layer edge detection and depth or-
dering, and present our post-processing procedure as follows.
The motion models of the layers are determined by the subspace segmentation ap-
proach described in section 2, while the layer boundaries and the layers of the oc-
cluded regions are undetermined. The problem we face here is how to determine the
layer boundaries and infer the occlusion relationships. In order to do that, we will
consider the intensity structures of each frame, the relevant occlusion region map
(obtained by [15]) and the relevant boundary map of the initial motion layers (obtained