Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 31

is taken by a camera undergoing translation from the left to the right. Our goal is
to determine the boundaries of the motion layers, and find the layer order over two
successive frames. Our approach found out two motion layers, the tree trunk and
garden background.
Fig.3 gives the segmentation of the affine model using the subspace segmenta-
tion approach described in section 2. It can be noted that the occlusion regions
from the optical flow fields are crude, and contain many spurious small regions.
The red arrows illustrate the possible occlusion regions in the successive frame 1
and 2. We show the results of motion segmentation of frame 1 in Fig.3(1-3). Note
that the occlusion regions in Fig.3(2-3), which are not the layer boundaries, are the
interim areas between the foreground and background. It is impossible to deter-
mine the depth ordering using the obtained motion layers before determining the
layers of the occluded regions. In addition, the resulting layer boundaries are also
unreliable. Similar to the occluded regions, there are many small and isolated
spurious regions on the obtained layers. We need to refine the layer boundaries
and find out the layer order.
frame 1 frame 2
1. Occlusion regions 2. 3.
Fig. 3 Motion segmentation results by the subspace segmentation approach only. 1) The
occlusion regions from the optical flow field between frame 1 and 2; 2) and 3) are the
segmentation results only using the subspace segmentation approach described in section 2.
Fig.4 gives the segmentation results of the subspace segmentation approach fol-
lowed by the post-processing procedure described in section 3. The boundaries of
motion layers can go across the occluded regions and converge to the desired
locations. But we can also note that a patch of ground is classified as the fore-
ground as shown in Fig.4(1). This is due to the fact that the motion variance of this
patch between the successive frames 1 and 2 is close to that of the “tree”, away
from the background. If going through over multiple frames, the motion of this
patch should be distinguished from the “tree”, since the motion of the ground is
prone to be modeled by a single affine model. The intensity edge map of frame 1