Tuesday February 24, 2009
Hamerschlag Hall D-210
Carnegie Mellon University
The goal of stereo vision is to reconstruct disparity map from a stereo image pair. Most existing stereo algorithms can be classified into either local or global approaches. Generally speaking, local algorithms are computationally less expensive, while global algorithms produce a higher quality disparity map. Recent research shows that local approaches based on carefully designed cost aggregation strategies can outperform many global approaches.
Among local aggregation approaches, adaptive-weight window produces the best quality disparity map under real-time constraints but is relatively slower. We propose a fast adaptive-weight aggregation method based on exponential step information propagation. The idea is to propagate information from long distance pixels within a few iterations.
In this talk, I will also present efficient implementation techniques on a GPU platform, which achieve 10.5x speed up over a straightforward implementation. Experimental results show that our technique is Pareto-optimal among existing real-time or near real-time stereo algorithms along the accuracy-speed trade-off space.
Wei Yu is a third year PhD student at Carnegie Mellon (ECE), advised by James C. Hoe and Tsuhan Chen. Her research interests include acceleration of visual computing on multi-core platforms.