CS348B Final Project
Group: Matthew Everett and Jeffrey Mancuso
For our final project we chose to render Escher's Reptiles drawing. The primary goal of this project was to create the most realistic image possible while remaining true to the original work.
Below are the final images we rendered at 100 samples per pixel. The above black and white image was converted to grayscale, gamma corrected and had a slight increase in aspect ratio from the final rendering on the left.
Original perspective Side view
In order to create a realistic image, as well as make it as close to Escher's drawing as possible, we implemented a number of enhancements to LRT.
Vanishing points in photographs can be distracting because the perspective of the image is correctly only when looking at the center of projection. Escher drew "Reptiles" with an overhead viewpoint looking down to the ground plane. Objects that are perpedicular to the ground, such as the shot glass and the bottle, are completely vertical in the drawing. A standard perspective transformation would normally create a distortion so that the bottle and the shot glass would be bent towards the right. In order to fix this, we implemented a tilt-shift lens - common to view cameras and high end 35mm systems.
The perceived distortion of vanishing points is evident in the following image:
The tilt-shift lens solves this problem by effectively rotating the film plane to be perpendicular to the ground plane. The vanishing points then disappear, as does the perceived distortion. We simulated this effect by allowing graphic artists to add a "pitch_angle" parameter to perspective transformations that rotates the view of the camera without changing the orientation of the film plane.
This portion of the project was implemented by changing the parameters passed to the Frustum transformation. The center of the image is rotated about the eye point by pitch_angle degrees. The extent of the frustum in the y direction is then expanded to maintain the original field of view. The distortion of the previous image is eliminated in the corrected version:
The cones now all point upward, as would be expected by someone looking directly at one of the cones.
Bump Mapping & Multi-Texturing Enhancements
In order to make the image as realistic as possible we implemented bump mapping. A bitmap is used to represent the heights of bump offsets. These bumps can be amplified or even inverted using a bumptextureamount variable. The bump mapping code was placed in the shade context so that it is accessible from all shaders and to make it apparent during recursive reflections. Multi-texturing is also in place to allow for more realistic objects. Nearly every object in the scene was textured in one way or another. Without this, many things had a plastic feel to them that detracted from the scene.
Bump Mapping and Multiple Textures
Since we were unable to perfectly model the reptiles, we decided to enhance their appearance with an Iridescent shader. Without implementing an actual simulation of iridescent effects, we developed a multicolored, anisotropic, view-dependent surface shader. It gives the impression of having a wavelength-dependent surface. The surface was implemented by combining the information from the [s, t] coordinates for vertices with the cosine of the half-angle to form a series of swirling spots on the surface. These spots have a center of color as well as a center of visibility. The spots are generated randomly using a repeatable random number algorithm. Areas not covered by these swirling spots are filled in by a separate diffuse color. The following image demonstrates the final surface:
We also modified lrt to run on multiple machines simultaneously through the standard MPI interface. One node is designated the master node and collects the sample data from the other nodes. The other nodes are each assigned a block of samples to evaluate. When these nodes are done, they send them via MPI to the master node, which adds them to the image and writes the image once the samples are collected. We achieved qualitatively faster rendering speeds on clusters than we would have otherwise. An quantitative evaluation of the speedup remains to be done.
Light Field Viewer
We modified lrt to produce "light fields," which are collections of samples rather than images. These collections can be used to simulate several interesting effects. Our light field renderer/viewer has the following features:
The interface has several controls:
These images are rendered from a light field consisting of a 16x16 array of 100x100 pixel images. Note the supersampling in the original images that allows the light field viewer to generate high-quality images from these small initial images. Gray pixels indicate points where sampling was unsuccessful.
Depth of Field
LRT & Reptiles Scene