Mapping onto a sphere

Figure 7.24: (a) The photophere is texture-mapped onto the interior of a sphere that is modeled as a triangular mesh. (b) A photosphere stored as a cube of six images can be quickly mapped to the sphere with relatively small loss of resolution; a cross section is shown here.
\begin{figure}\begin{center}
\begin{tabular}{cc}
\psfig{file=figs/spheremesh.ps,...
...ap.eps,width=2.7truein} \\
(a) & (b) \\
\end{tabular}\end{center}
\end{figure}

The well-known map projection problem from cartography would be confronted to map the photosphere onto a screen; however, this does not arise when rendering a photosphere in VR because it is mapped directly onto a sphere in the virtual world. The sphere of all possible viewing directions maps to the virtual-world sphere without distortions. To directly use texture mapping techniques, the virtual-world sphere can be approximated by uniform triangles, as shown in Figure 7.24(a). The photosphere itself should be stored in a way that does not degrade its resolution in some places. We cannot simply use latitude and longitude coordinates to index the pixels because the difference in resolution between the poles and the equator would be too large. We could use coordinates that are similar to the way quaternions cover the sphere by using indices $ (a,b,c)$ and requiring that $ a^2 + b^2 + c^2 = 1$; however, the structure of neighboring pixels (up, down, left, and right) is not clear. A simple and efficient compromise is to represent the photosphere as six square images, each corresponding to the face of a cube. This is like a virtual version of a six-sided CAVE projection system. Each image can then be easily mapped onto the mesh with little loss in resolution, as shown in Figure 7.24(b).

Once the photosphere (or moviesphere) is rendered onto the virtual sphere, the rendering process is very similar to post-rendering image warp. The image presented to the user is shifted for the rotational cases that were described in Figure 7.19. In fact, the entire rasterization process could be performed only once, for the entire sphere, while the image rendered to the display is adjusted based on the viewing direction. Further optimizations could be made by even bypassing the mesh and directly forming the rasterized image from the captured images.

Steven M LaValle 2016-12-31