Graphics Reference
In-Depth Information
1 spatial sample
4 samples
1,024 samples
Actual Scene
1 temporal sample
1,024 samples
4 samples
Actual Scene
Figure 35.12: Top row: a fence made of black posts on a white background imaged with
increasing spatial resolution. Bottom row: a moving sphere imaged with increasing tempo-
ral resolution. Increasing the number of samples better captures the underlying scene, in
space or time. Here a regular sampling pattern is used for each.
sample, the car may be either present or absent. Increasing the temporal sampling
rate increases the ability to resolve the car. For a high temporal sampling rate the
image approaches that which would be captured by a real camera, where the car
is blurred across the entire frame.
One method for ameliorating temporal aliasing is to use a high-refresh rate
display and render once per refresh. Although not common today, there are pro-
duction 240 Hz displays. Simply rendering at 240 Hz provides four times the
temporal sampling rate of the common 60 Hz rendering rate. This does not solve
the temporal aliasing problem. It merely reduces its impact. A sufficiently fast car,
for example, will still flash into the center of the screen and disappear.
A more common alternative to a high refresh rate solves the flashing problem
and does not require a special display. One can explicitly integrate many tempo-
ral samples, producing rendered motion blur. Distribution
1
ray tracing [CPC84]
pioneered this approach, which has since been extended to rasterization. Here,
software is performing the integration that was performed by the eye under a high-
refresh display.
Integration over temporal samples does not necessarily give the same percep-
tion as observing a high refresh display or the real world, however. The reason is
that the eye is not a camera. In the absence of temporal integration, the observer's
eye can track the motion of an object in the scene at a high rate. For example,
the eye can rotate to keep an object moving across the display's field of view at
the same location in the eye's field of view. The resultant perception is that the
moving object is sharp and the background is blurred. If the same scene is shown
at a low frame rate that has been integrated over multiple temporal samples, the
moving object will be blurred and the background will be sharp. This might lead
one to the conclusion that motion blur cannot be rendered effectively without eye
1. Cook et al. originally called their technique “distributed” ray tracing because it
distributes samples across the sampling domain, including time. Today it is com-
monly called “distribution” ray tracing to distinguish it from processing distributed
across multiple computers. It is also called “stochastic” ray tracing since it is often
implemented using stochastic sampling, although technically, the decision to distribute
samples (especially eye-ray samples) is separate from the choice of sampling pattern.
