The problem of reconstructing
real points of view of an arbitrary 3D scene
and synthesizing virtualized
ones, from a limited set of 2D uncalibrated
views without resorting
to a 3D CAD model of the scene is referred to as
Video Spatialization. This
is called Video Spatialization rather than View
Interpolation, since the
3D information given by triplets of different points
of view of the scene (i.e
sequences of three different views) is taken into
account via the
trilinearity
theory.
An efficient
``mesh-oriented''
approach for real view regeneration
from two
neighboring ones was developed
and Image Mosaicking was studied as
a pre-
processing step to increase
limited overlapping areas between the initial input
data (2D uncalibrated images).
Video spatialization aims
at offering the possibility for a user to visualize the
scene from anywhere and
in any direction, just as he really was immersed in a
3D place. In this respect,
a step of analytical inferences to synthesize
the
coherent points of view
in relation with the user's motions and orientations
is needed.
These algorithms, based on
a set of three uncalibrated views, are called
``intra-triplets'' processings,
however lots of applications will use more than
three views: these processus
must be extended combining the synthesis method
from several triplets of
uncalibrated images with Image Mosaicking approxi-
mations in order to simulate
global motions through ``virtualized'' 3D scene.
Such extensions are called
``inter-triplets'' processings.
In addition to common applications
like virtual walks and visits, we use
our
approach in the context
of multipoint teleconferencing systems for very low
bit rate links (internet,
mobile communications). In the context of the
TRAIVI
project, Video Spatialization
is required to control background environment
and especially to maintain
the coherence between the virtualized user's point
of view of the meeting area
and his positions, orientations and motions during
the videoconference session. |
