Information Technology Reference
In-Depth Information
(a) (b)
Fig. 6
(a) Grouping in the horizontal direction. (b) Grouping using adjacent lenses.
4 3D Object Segmentation
Research on depth extraction from multi-view imaging systems has been exten-
sive. However, the depth extraction from 3D holoscopic imaging systems is still in
its infancy. The first reported work is that of Manolache
et. al.
[36], where the
point-spread function of the optical recording is used to describe the associated in-
tegral image system and the depth estimation task is tackled as an inverse prob-
lem. In the practical case, the image inverse problem proves to be ill posed and the
discrete correspondents are ill conditioned due to the inherent loss of information
associated with the model in the direct process. Therefore, the method can only be
applied on simulation using numerical data [37, 41].
A practical approach for obtaining depth by viewpoint image extraction and dis-
parity analysis was explored and presented by Wu,
et. al
. [37]. The viewpoint im-
age was formed by sampling pixels from different micro-images rather than a
macro block of pixels corresponding to a microlens unit. Each “viewpoint image”
presented a two-dimensional (2D) parallel recording of the 3D scene from a
particular direction. Figure 7 graphically illustrates how the viewpoint images are
extracted.
Figure 8 shows one example of unidirectional 3D holoscopic image and the ex-
tracted viewpoint images. The number of pixels of the formed viewpoint image
(along horizontal direction) will depend on the size of the 3D holoscopic image. In
the typical case described in figure 8, where the pitch size of the microlens sheet is
600 µm and the image has a size of 12cm, there will be 200 pixels in one formed
viewpoint image along the horizontal direction. Assuming that the intensity distri-
bution has been sampled so that each micro-image comprises 8 pixels in the hori-
zontal direction. This results in eight viewpoint images being extracted as shown
in figure 8b.
