Image-Based Rendering - The Magic of Computer Graphics - page 137

Viewpoint of camera 2

Viewpoint of camera 1

Viewpoint of
virtual camera

b 1

b 2

A B C

Model

Figure 5.11 View-dependent texture mapping (VDTM). At an arbitrary viewpoint (indicated by the ar-
row) the camera sees point B , which corresponds to points in images captured by cameras 1
and 2. The value of the texture at B is a weighted average of the values in captured images.
The weights are determined by the deviations b 1 and b 2 from the viewing angles of the
captured images. (After [Debevec et al. 96].)

Stage 2. View-dependent texture mapping (VDTM): Mapping the
captured images to the recovered model. The ultimate goal is to re-
render the model from an arbitrary viewpoint. It is assumed that the images are
captured under approximately the same lighting conditions. However, the sur-
faces of the object being modeled are not likely to exhibit purely Lambertian
reflectance, so their appearance depends on the view direction. The authors ap-
ply a view-dependent texture mapping approach to account for this. Figure 5.11
illustrates the concept. The pixel value seen by a virtual camera is interpolated
from the corresponding pixels in nearby captured images as a weighted average.
The weights are inversely proportional to the angles between the virtual view di-
rection and the corresponding capture view directions. Figure 5.9(d) shows the
model renderedwith the view-dependent texture applied; capturing an image from
this high viewpoint would be difficult.

Stage 3. Model-based stereo: Adding detail to the 3D model us-
ing image warping The basic model constructed in the first stage is only a
rough approximation to the underlying architecture model. As explained above,
this simplifies fitting the parameters. However, the finer details of the model are
important—rendering using the rough model cannot be expected to look right.
For example, the rough model of the bell tower uses a single block for the sec-
tion containing the bells: the arches and pillars and the inside elements are not
included in the rough model.