Project Overview

The Cube project is one of the primary research efforts in the Visualization Laboratory of the Computer Science Department at the State University of New York at Stony Brook. It was started in 1982 by project director Prof. Arie Kaufman. The primary goal is the design and implementation of a special-purpose volume rendering accelerator for the real-time visualizaton of high-resolution volumetric datasets.

Support for this project has been provided by the National Science Foundation, Hewlett Packard Laboratories, Japan Radio Corporation and Mitsubishi Electric (MELCO) and Mitsubishi Electric Research Lab (MERL).

People

Arie Kaufman – Project Director
Patrick Tonra – Engineer
Hanspeter Pfister – PhD. Student (Alumnus)
Ingmar Bitter – PhD. Student (Alumnus)
Baoquan Chen – PhD. Student (Alumnus)
Frank Dachille – PhD. Student (Alumnus)
Kevin Kreeger – PhD. Student (Alumnus)
Michael Meissner – M.S. Student (Alumnus)
Urs Kanus – M.S. Student (Alumnus)
Mohamed AbdelSadek – Undergraduate Student (Alumnus)
Kwasi Mireku – Undergraduate Student (Alumnus)
Fayyaz Younas – Undergraduate Student (Alumnus)

Real-Time Hardware for Volume Visualization

Motivation

Volume rendering is a key technology for the interpretation of the large amounts of 3D scalar data generated by acquisition devices such as biomedical scanners or by supercomputer simulations. Of particular importance for the exploration and understanding of the data are sub-second display rates and instantaneous visual feedback during the change of rendering parameters. To create the illusion of smooth dynamics, the image must be updated at true real-time rates.

Objectives

The main goal of our research is to develop hardware architectures for real-time volume rendering of high-resolution datasets. Our research has the following design objectives based on what we believe to be important features of a real-time volume rendering system:

Real-Time Frame Rates: To create the illusion of smooth motion, the image must be updated a minimum of 24 times per second. We aim at achieving projection rates of 30 frames per second.
4D Visualization: The architecture has to allow to directly view changes of the 3D data over time for 4D (spatial-temporal) visualization, such as in real-time 3D ultrasonography, micro-tomography, or confocal microscopy. This requires the real-time input of volumetric data without pre-computations. The overall latency of the system should be no more than one frame time.
High-Resolution Datasets: The architecture has to be able to visualize dataset resolutions of 512 ³ voxels or higher in real-time.
Scalability: The design should be modular, and the performance should ideally scale almost linearly in the number of modules.
High Image Quality: The images must be of high quality, including surface shading, depth cues, and the provision of transparency. Special care has to be taken to avoid image artifacts such as spatial or temporal aliasing.
Flexibility: The algorithm and hardware should be flexible enough to allow for the interactive change of parameters such as shading, data segmentation, and projection modes.

Why Special-Purpose Hardware?

Current general-purpose systems fall short of achieving these goals. The high computational requirements of traditional computer graphics led to the development of special-purpose graphics engines, primarily for polygon rendering. Similarly, the special needs of volume rendering, where an image must be computed rapidly and repeatedly from a volume dataset, lends itself to the development of special-purpose volume rendering architectures. A dedicated accelerator, which separates volume rendering from general-purpose computing, seems to be best suited to provide real-time volume rendering on standard deskside or desktop computers.

Volume rendering hardware may also be used to directly view changes of the 3D data over time for 4D (spatial-temporal) visualization, such as in real-time 3D ultrasonography, micro-tomography, or confocal microscopy. This may lead to the direct integration of volume visualization hardware with real-time acquisition devices, in much the same way as fast signal processing hardware became part of today’s scanning devices.

Finally, the choice of whether one adopts a general-purpose or a special-purpose solution to volume rendering depends upon the circumstances. If maximum flexibility is required, general-purpose appears to be the best way to proceed. However, an important feature of graphics accelerators is that they are integrated into a much larger environment where software can shape the form of input and output data, thereby providing the additional flexibility that is needed. A good example is the relationship between the needs of conventional computer graphics and special-purpose graphics hardware. Nobody would dispute the necessity for polygon graphics acceleration despite its obvious limitations. We are making the exact same argument for volume rendering architectures.

Cube-4 – A Scaleable Real-Time Architecture

Our latest architecture, Cube-4, performs arbitrary parallel and perspective projections of high-resolution datasets at true real-time frame rates. The performance is data and classification independent and can be achieved at a fraction of the cost of a multiprocessor computer. Cube-4 uses accurate 3D interpolation and high-quality surface normal estimation without any pre-computation or additional storage. Consequently, Cube-4 is also appropriate for 4D visualization as an embedded volume visualization hardware system in real-time acquisition devices.

The Cube-4 architecture performance grows proportionally with increasing number of memory and processing units, ultimately limited by memory speeds. Possible hardware implementations of Cube-4 for 30 frames per second rates range from an inexpensive PCI board accelerator for 256 ³ datasets, to a workstation accelerator board for 512 ³ datasets, to a high-end visualization server for 1024 ³ or higher resolutions. The cost-performance ratio of Cube-4 is several orders of magnitude better than existing solutions.