Wineglass
CS348B Final Project
Jeff Brasket, Michael deLeon
Out final rendered image (top left) vs. our target photograph
... click
here for 800x800
resolution
Motivation
For our final project, we wanted to experiment with an image
that would showcase caustics as portrayed by photon mapping. Searching the
internet, we started thinking about wineglasses. We found an interesting
mix of images from wineries in outdoor settings and a fair number of unrealistic
wineglass images scattered across the web from various past projects. We
decided we wanted to make a very realistic looking wineglass illuminated by an
outdoor scene. See our proposal for
more about our inspiration.
We wanted to include an outdoor photograph for our backdrop, so our wineglass
should have to look very convinving to stand up to this comparison.
Photon Mapping
The most challenging part of our project involved
implementing the photon mapping algorithm as explained in Henrik Wann Jenson's
excellent book, Realistic Image Synthesis Using Photon Mapping. After
developing a general global photon map and a way of "painting" dots to
demonstrate the placement of photons, we next worked on a separate caustic
photon map to increase the resolution of the caustics generated by the glass
that would be the centerpiece of our image. We did render a few Cornell
Box-like test images to make sure our algorithm was working correctly. We
also ended up finding a fiar amount of bugs in the LRT Path Tracing integrator
in the process of exploring diffuse reflections with this box. We later
realized that these diffuse reflections would add little to our particualr
schene, and so built a modified Whittig ray tracer with photon mapping activated
under the Monte Carlo heading to enable us to call either from RIB files.
Our final caustic map can be targeted using one or two ciircles specified in
the photon mapping code. Samples are regularly generated over this area,
bounced around according to a russian roulette system, and then stored in
Jenson's photon map data structure. For the caustic photon map, we ony
store photons from this target that hit specular surfaces on the first
bounce. This gives us an excellent sample of the caustic genreating
photons. Eventually, we also used the cognac glass data from Jensen's book
make sure our glass model and photon map systems were working correctly.
testing our photon model with the cognac data ...
NURBS
We decided to model our wineglass in Maya and exported it to the
RIB format. Past projects told us this would be a challenge, but
possible. We had the most luck exporting revovled surfaces to the RIB
format using a standard Maya plugin and then using only the NuCurve information
in our RIB files. We seperated the wine surface and wine/glass interfaces
into distinct revovled surfaces, and even seperated the base and glass upper lip
so that we could improve resolution in these areas when refining the NURBS
surfaces. The NURBS support in LRT seemed to work pretty well, though we
added the ability to interpolate normals, gererating a normal for each vertex
directly from the NURBS curves and storing them for interpolation during the
resulting triangle mesh refinement. This gave us a very smooth, realistic
glass appearance. We tried several iterations of our wineglass model
before yielding one that looked very close to our example glass.
Earlier versions .. a
non-interpolated, fine mesh glass (left) and early table texture attempts with
thick-walled glass model.
Modeling Glass
We started our glass simulation by implementing the Fresnel equation for
glass reflection and transparency as described in lecture. We experimented
carefully with indexes of reflection to model the glass and wine differently,
and used the ratio of the two to get the interface right with LRT. We also
experimented briefly with a normal map to simulate minor imperfectiosn in glass
thickness and smoothness, but did not have time to get this looking realistic.
Backgrounds
We took several reference images using a diginal camera, including target
photos of wineglasses under various sunlight conditions. We also took
extra background images to use with our rendered image. We even used a
silver ball to attempt environment mapping. With a little creative
placement of these images all around our subject and careful lighting placement,
we were able to effective recreate our scene conditions under all transmission
and refelctions directions in our glass. We used a silver ball and glass
sphere as test objects to make sure all directions were decently represented in
our test image. We also took time to accurately model a table surface for
our glass with a realistic texture pulled from a Pergo web page and dailed in
for proper illumination.
Results
We were very happy with our results. Many folks have
trouble distinguising between our image and the photo that inspired it. We
used a final photon map of about 200,000 photons and a rendering sampling of 16
rays per pixel to eliminate aliasing. Although our methods were fairly
straightforward, we think we showed that very realistic images are possible
using basic but well implemented ray tracing techniques.
An
image with red wine and a different viewing angle