Parallel Computer Vision

This is a "sampler" page of the Robotics Institute at Carnegie Mellon University.

This project applies advanced, low-latency supercomputers to problems in computer vision.

Early work led to the use of the Carnegie Mellon Warp machine in the Navlab robot vehicle. A Warp machine was mounted in Navlab and used for various tasks, including road following using color-based image segmentation, and also using the ALVINN neural-network system.

The ALVINN work was particularly successful, because an important factor in developing the road following neural network was the availability of Warp to perform back propagation on large neural networks.

Current Work

Managing large data structures. We are studing how algorithms that manipulate large data structures can be mapped efficiently onto a distribute memory parallel computer.

Efficiently moving data structures. How can the most efficient of the many possible communications patterns for reorganizing data structures within a parallel computer be generated automatically?

Developing parallel vision software. We are implementing a parallel version of the ISO standard Programmer's Imaging Kernel System. This implementation is based on Adapt, which has been mapped onto the Carnegie Mellon-Intel Corporation iWarp computer and the Intel Paragon. An implementation for the IBM SP1 is planned.

Real-time stereo vision. We have implemented the fastest stereo vision system ever demonstrated. It uses Kanade-Okutomi multi-baseline stereo and operates at 15 Hz on a 64-cell iWarp, turning three 240x256 input images into a 240x256x16 depth image. This system was recently demonstrated at Supercomputing '93, as shown below.

We have built an elegant four-camera multibaseline stereo system, as illustrated below, which is capable of capturing images with less than 1 mm error over a range of depths of meters. This uses an active vision technique that allows create flexibility, accuracy, and high-speed video capture. All processing is done in Fx, a portable variant of high performance Fortran.

Future Work

Our current interest is in building computer vision systems that use parallel computers to achieve performance never previously demonstrated:

We are studying issues in higher level parallel vision in the context of an automatic tool for generating parallel object recognition programs, based on the Vision Algorithm Compiler.

We are working to combine the stereo vision system with high-quality graphical output in the High Performance Computing Graphics project.

We are applying parallel image processing and computer vision techniques to problems in biology as part of the Automated Interactive Microscope (AIM) project.

Acknowledgments

Work on managing large data structures: G. Gusciora Ph.D. Thesis. Webb, H. T. Kung advisors. Completed May 1994.

Work on efficiently moving data structures: D. Smith Ph.D. Thesis. Webb, H. T. Kung, advisors. Expected 1994.

Implementations of PIKS: Intel work done by M. MacPherson. IBM SP2 implementation to be done by G. Gusciora at the Maui High Performance Computing Center.

Work on real-time computer vision: Webb, T. Warfel, S. B. Kange.

Work on stereo: S. B. Kang, Webb, C. Lawrence Zitnick, Takeo Kanade.

Work on parallel object recognition: M. Wheeler, K. Ikeuchi, Webb.

Work on high performance computing graphics: P. Heckbert, Webb.

Work on AIM project: P. Janardhan, K. Ikeuchi, Webb.