Recent advances in multimedia technology and virtual reality applications have
created a wide range of applications where computer generated 3-D environments
are desirable. Applications include architecture visualization, virtual
television studios, virtual presence for video communications, and general
``virtual reality" applications. In all of these, real objects must be
scanned and 3-D models must be generated automatically with low cost.
The authors wish to reconstruct 3-D models of objects from stereoscopic image
sequences. The system may be either data driven or may interpret the scene
based on explicit knowledge about the objects. The system consists of three
modules:
Stereoscopic image analysis: Extracts 3-D features from the image
sequence
3-D reconstruction: Solves the problem of view point estimation
and data fusion from multiple view points
Scene interpretation: Controls
the modeling and maintains the modeling and scene knowledge
Given a stereoscopic pair of images of a piecewise planar object, this paper
presents:
A method for computing local surface orientation and a segmentation of
the object from quantized disparity maps that are computed by stereoscopic
correspondence analysis
A method for converting this segmentation into a set of planar surfaces
that make up a 3-D polygonal model of the object; this model is made
photorealistic by fitting planar polygons to the surfaces and mapping original
image texture to the model
The method presented in this paper still requires the user to specify the
segmentation threshold and classify which surfaces are relevant for the
modeling and which can be discarded during the approximation. Future work
includes further automation of this process during scene interpretation.