projects

Artistic Multiprojection Rendering

Maneesh Agrawala	Computer Graphics Laboratory, Stanford University
Denis Zorin	Media Research Laboratory, New York University
Tamara Munzner	Computer Graphics Laboratory,

In composing hand-drawn images of 3D scenes, artists often alter the projection for each object in the scene independently, thereby generating multiprojection images. We developed an interactive tool for creating such multiprojection images and animations, consisting of two parts: a multiprojection rendering algorithm and an interactive interface for attaching local cameras to the scene geometry.

Artists create such multiprojection images for several reasons, including: expressing a mood, feeling or idea; improving the representation or comprehensibility of the scene; and visualizing information about the spatial relationships and structure of the scene. Multiple projections could similarly enhance computer-generated images and animations, but simple and efficient methods for multiprojection rendering have not been available. We consider a variety of examples from tradtional art that use multiple projections.

Artistic Expression

In Giorgio de Chirico's The Mystery and Melancholy of a Street, figure 1, the buildings, the van and the ground plane all have different viewpoints. A melancholy aura is created by the unusual arrangement of the objects which results in an incongruous spatial system. Despite the large disparities between projections, the overall impression of a three-dimensional space remains intact.

Cezanne similarly incorporates multiple viewpoints in Still Life with Fruit Basket, figure 2. These viewing distortions generate tension between different planes in the image; the distortions flatten some regions of the picture, while enhancing depth in other regions. He explains that the inconsistencies in projection generate an ``emotional nonrealistic illusion of space.''

Representation

Certain viewpoints are better than others for comprehending the overall geometric shape of an object. By using multiple viewpoints the artists can present the best views of all objects in the scene. A single viewpoint is often inadequate for large format pictures like murals, frescoes, or billboards. Such images are often placed above standard eye-level and are seen from a much wider range of viewpoints than smaller format pictures. For these reasons, many large format images are created with multiple projections. Paolo Uccelo's fresco of Sir John Hawkwood is a well-known large format multiprojection example. He used two projections, one for the base and one for the horse and rider, to produce the best view of both elements, as described in figure 3.

A well known problem with wide-angle, perspective projections is that curved objects located near the edges of the viewing frustum, close to the image plane, can appear unnaturally stretched and distorted. The most common technique for decreasing this distortion is to alter the projection for every object to provide a perceptually ``correct'' view of each one. The foreground human figures in Raphael's School of Athens, figure 4, are inconsistent with the strong central perspective projection of the background architecture. The inconsistencies improve the comprehensibility of the figures and make them easier to recognize.

Visualization

It is possible to create an artificial sense of perspective using multiple axonometric or multiple oblique projections. The ``trick'' is to orient the receding parallels of each object towards some pre-chosen vanishing point. In the illustration of 53rd Street (figure 6) the buildings are drawn from above in oblique projection. The receding parallels for each building point towards a central vanishing point, thereby creating the illusion of perspective.

Our Approach

We present interactive methods for creating multiprojection images and animations. The main technical contributions of our work are new algorithms designed for:

Resolving Visibility: In the multiprojection setting there is no uniquely defined solution to the visibility problem. However, in many cases the user wishes to maintain the visibility ordering of a {\em master camera} while using different {\em local cameras} to introduce shape distortions to individual objects. Based on this insight, we propose an algorithm that automatically resolves visibility for most practical cases and allows user adjustments when the automatically computed visibility is not satisfactory.

Constraining Cameras: We suggest a simple and intuitive set of camera constraints allowing the user to choose appropriate projections for a variety of artistic effects. These constraints are particularly effective when initially placing cameras and when animating the scene.

Interactive Rendering: We leverage multipass hardware rendering to achieve interactive rendering rates. The user can immediately see how changing the parameters of any camera or moving any object will affect the final image.

Images rendered using our system appear in figures 7-10. We have also produced several animations using our system.

Relevant Links

Eurographics Rendering Workshop 2000 paper

QuickTime video segments. Although these individually downloadable segments are half-size (320x240 pixels) instead of full video resolution (640x480 pixels), the files are still quite large and are not suitable for download over a slow connection.

• mpr.00.mov 1:19 min:sec [11.6MB]
• mpr.01.mov 0:52 min:sec [7.5 MB]
• mpr.02.mov 0:23 min:sec [2.2 MB]
• mpr.03.mov 0:59 min:sec [10.2 MB]
• mpr.04.mov 1:13 min:sec [4.5 MB]
• mpr.05.mov 0:32 min:sec [3.4 MB]

Our reconstruction of de Chirico's Mystery and Melancholy of a Steet.

A multiprojection still life created with our system