Apply GPCA to Motion Segmentation - Intelligent Video Event Analysis and Understanding - page 32

1. foreground 2. background
3. Intensity edge map 4. Occluding edges (red)
Fig. 4 Refined and layered motion segmentation. 1) and 2) are the segmentation results; 3)
The intensity edges of the image; 4) Occluded region and occluding edges.
is obtained by the Canny edge detector, and also shown in Fig.4(3). It can be ob-
served that the boundaries of motion layers are involved in the intensity edge map,
e.g. the red arrows illustrate the corresponding edges between the layer boundaries
and the intensity edges of the image. Moreover, we also show the occlusion edges
(red) in Fig.4(4). Partial occluding edges are not involved in the initial occlusion
regions. But, the profiles of these edge points overlap with the occlusion regions.
These points can thus be joined with the occluding edges. Additionally, it can be
noted that the layer boundaries involve the occluding edges, but the layer bounda-
ries are not always the occluding edges. Locating the occluding edges can help us
find the depth ordering.
Susie Calling
This sequence presents a hand holding a phone while the head is rising slightly.
The image resolution is 170 μ 120 pixels. It can be observed that the region of the
phone is enwrapped by the head region. Our segmentation approach aims at sepa-
rating the phone region from the head region. The segmentation results are shown
in Fig.5. The region of phone is in front of the regions of head and background.
The background region is behind the head region.
Due to the rich texture of the hair, the segmented head region contains many
small holes, particularly in the hair area. It is difficult to determine the boundaries
of the hair. For example, in Fig.5(2), a patch of hair image is incorrectly classified
into the group of the phone region. The post-processing could not merge this patch
into the head region either, as shown in Fig.5(4-6). This is because the detected
layer edge segments goes through the occluded regions as shown in Fig.5(7).