Ray tracing is ideal for showing dynamic effects such as specular highlights and shadows. Dynamic objects are more difficult to incorporate into a ray tracer than into a z-buffer algorithm as current acceleration schemes are not dynamic [11]. Our current workaround is to keep dynamic objects outside the acceleration scheme and check them individually for each ray. Obviously this only works for limited numbers of dynamic objects. In Figure 2 we show a static image from a set of bouncing balls using the soft shadow approximation.
Computer-aided design usually uses both
curved surfaces and non-diffuse objects, such as a windshield
made from glass. Ray tracing can
render curved surfaces directly,
making it ideal for spline models. The ability
to calculate accurate reflections across the surface make is possible
to evaluate the smoothness and curvature of the models for aesthetic
purposes. A sample of a directly ray traced spline primitives
is shown in Figure 14. We have run on several models
containing 20-2000 individual patches with runtimes ranging from
1-20 fps at 512 by 512 pixels on 60 processors.

Ray tracing time is sub-linear with respect to model size. This allows us to interact with very large models. One area that creates large models is scientific visualization. In Figure 15 we show a visualization of a stress simulation. Each node in the simulation is represented by a sphere. There are 35 million spheres in this model. Unlike conventional rendering systems, the high depth complexity has very little effect on the rendering times. Another area that can create complex models is architectural design. The model in Figure 16 contains roughly 75,000 polygons and a spline teapot. An area we would like to explore is the use of interactive ray tracing for walk throughs of globally illuminated static environments, where the illumination information has been computed in advance by such techniques as radiosity or density estimation. Usually specular and transparent effects are missing from such walk throughs. In addition, we should be able to easily allow higher order reconstruction of the solution. Also, we could greatly reduce polygon count if radiance lookup evaluates the mesh instead of representing each mesh element as a separate polygon.

