Abstract:
|
English: In the last few decades a number of methods have been proposed for the
acquisition of 3D models from real objects, including 3D scanning, 3D point-
based photography and image-based modeling, among others. These tech-
niques provide users with highly-detailed models, especially for simple and
convex objects, which can be sampled from all the view directions surround-
ing the object. However, the detailed acquisition of very large models of
natural environments is a challenging problem for several reasons, including
inaccessibility (e.g. forest area), insu cient lighting (shadowed areas) and
presence of highly self-similar structures. For example, point-based models
of outdoor environments often combine highly-detailed areas with low-detail
or even unsampled areas. In this work we propose a reconstruction method
for enhancing and enriching the geometry and appearance of point-based
models of natural environments. Although our technique is orthogonal to
the original acquisition technique, we mostly focus on enhancing models ac-
quired through image-based modeling techniques such as multi-view stereo.
Our method uses information from highly-detailed areas to ll unsampled or
poorly sampled areas. Since most components in natural environments, and
in particular the terrain, are fractal by nature, we propose to estimate the
fractal dimension of the surface in highly-detailed areas and to use fractal
generation techniques to both improve the overall appearance by increasing
the amount of detail and ll-in undersampled areas with new geometry hav-
ing similar fractal characteristics.
Our method starts by computing a triangle mesh approximating the input
point-based model. The main purpose of this step is to create a triangle
mesh lling the holes corresponding to poorly sampled regions of the input
model. Then, we subdivide the mesh by using a fractal perturbation scheme until the model has the desired level of detail. The output of our algorithm
is a triangle mesh which approximates the input point-based model but has
much more detail and better appearance. Although the resulting model is
not a faithful copy of the original environment, it provides a plausible visu-
alization with much higher quality and does not su er from undersampled
areas. |