Real-Time Analysis of Human Body Parts and Gesture-Activity Recognition in 3D - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 136

from prior information. In Yang et al. (2002), the authors use a graphics
hardware that effectively combines a plane-sweeping algorithm with view
synthesis for real-time, 3D scene acquisition.
Rendering
Rendering is the process of producing realistic 3D images. The rendering issues
are related to the interaction between light and surface, the intersection of
viewing rays and objects sampling of the scene and displaying techniques. There
are four main rendering methods used in visualization, i.e., ray tracing, volume
rendering, radiosity and polygon rendering (Crockett, 1997). Due to the high
computational requirements of traditional computer graphics, general purpose
computers are not efficient in rendering applications. Consequently, special-
purpose graphics engines are developed, primarily for polygon rendering.
Similarly, special-purpose volume rendering architectures are developed to meet
the special needs of volume rendering in order to compute rapidly and repeatedly
from a volume dataset. To provide real-time volume rendering on standard
computers, volume rendering is separated from general-purpose computing by
using a dedicated accelerator. Another approach is to use volume visualization
hardware that can be integrated with real-time acquisition devices.
3D reconstruction for image-based rendering is still an open research area. The
visual hull concept is introduced to describe the maximal volume that reproduces
the silhouettes of an object. In Matusik et al. (2000), an on-line, image-based
approach is described to compute and shade visual hulls from silhouette image
data. The maximal volume is constructed from all possible silhouettes. Compu-
tational complexity is reduced and a constant rendering cost per rendered pixel
is achieved. In Matusik et al. (2001), new algorithms are proposed to render
visual hulls in real-time. Unlike voxel or sampled approaches, an exact polyhedral
representation is computed for the visual hull directly from the silhouettes.
Several other methods are proposed for real-time rendering. Volume carving is
a common method used to convert silhouette contours into visual hulls by
removing unoccupied regions from an explicit volumetric representation. An-
other method is Constructive Solid Geometry rendering. To avoid the complexity
in computing the solid, ray tracing is used to render an object by defining a tree
of CSG operations. Although an image-based rendering method yields higher
realism, data acquisition and preprocessing requirements increase the complex-
ity.
In Goldlücke et al. (2002), a method based on warping and blending images
recorded from multiple synchronized video cameras is proposed to render