Graphics Reference
In-Depth Information
Light source
Viewpoint
Figure 2.5
Monte Carlo path tracing (MCPT) traces light paths starting from the viewpoint, continuing
in a random direction at each surface intersection until the light source is reached.
In this approach, sample points are chosen on the light source, and rays are
traced from each surface point to compute the visibility and the radiance. The
shadow value (irradiance) is then determined from the average value of these rays
in the usual Monte Carlo fashion. The work also extends ray tracing to the simu-
lation of real cameras. Focus and depth of field are rendered by tracing multiple
rays through the optics of a virtual camera. The paper also shows how sampling
can be distributed in time. The shutter of a real camera is open for a short, but
nonzero exposure time. Objects that move appreciably during the exposure appear
blurred. This
motion blur
comes from an integration over the time the shutter is
open. Giving rays a time value and allowing the scene objects to depend on time
allows motion blur to be computed by Monte Carlo integration.
The distributed ray-tracing work of Cook, Porter, and Carpenter was a signif-
icant step in realistic rendering, but the number of samples needed is prohibitive.
Furthermore, it does not produce a full GI solution: it only accounts for the
“blurred” phenomena of glossy reflection, soft shadows, and focus and motion
blur. The basic distributed ray-tracing algorithm does not account for diffuse
interreflection. Kajiya's “The Rendering Equation” paper [Kajiya 86] extends
distributed ray tracing to solve the GI problem using Monte Carlo path tracing.
Kajiya's approach is essentially the MCPT algorithm mentioned in Section 2.3.
At every ray-surface intersection a secondary ray is shot in a random direction;
at the surface that ray hits, another random direction is chosen; and so on, until a
light source is hit (see
Figure 2.5
)
. By sending enough rays from the viewpoint
or camera through each pixel (and averaging the results appropriately) essentially
all light paths can be captured.
In practice, efficiency can be improved by computing the direct lighting sep-
arately at each intersection point
x
on the path by directly sampling rays to the
light. In this case, a different path termination criterion is usually employed, such
as stopping at a prescribed depth. This approach or something similar actually
