Graphics Reference
In-Depth Information
The 3D eye position is then mirrored toward the screen, resulting in the correct
virtual viewpoint that is needed to restore the eye contact between the system users.
The two screens are placed in a common coordinate space, as if they were pasted
against each other. Hence, this creates the immersive effect of a virtual window into
the world of the other participant.
6.3.6 Networking
Our prototype system sends the eye coordinates over the network, and therefore
the requested image
I
v
can be computed locally at the peer that captures the rele-
vant images. These cross computations bring the required network communication
to a minimum, by avoiding the transfer of
N
input images. The total peer-to-peer
communication thereby exists out of the synthesized images and the eye coordinates.
6.4 Complexity Control
The previously discussed plane sweeping method is an efficient method to create
novel viewpoints. Nevertheless, we can increase performance even more by reduc-
ing the number of depth hypotheses. We propose 2 methods: an adaptive uniform
plane distribution method where the nearest and farthest depth values are adapted to
the scene, and an adaptive nonuniform plane distribution method, where the depth
planes themselves are redistributed in space to move computational power to the
places where there are actually objects. Bothmethods have other applications besides
videoconferencing and are discussed below.
6.4.1 Adaptive Uniform Plane Distribution
Most of the time, the head of a single person is visible in the camera views. To avoid
heavy constraints on the participant'smovement, a large depth range has to be scanned
to keep the complete head in the virtual view. This actually infers a lot of redundant
computations, since the head of the user only spans a small depth range. We therefore
propose to dynamically limit the effective depth range to
(similar
to [
10
,
23
]) through a movement analysis on the normalized depth map histogram.
This implicitly causes a quality increase of the plane sweep, as the probability of a
mismatch due to homogeneous texture regions is significantly reduced. Moreover, all
M
depth planes can be focused as
{
D
min
,...,
D
max
}
, which leverages
the dynamic range and thereby significantly increases the accuracy of the depth scan.
Three separate cases can be distinguished, as the user moves in front of the screen:
{
D
1
=
D
min
,...,
D
M
=
D
max
}
