Near Real-Time Stereo With Adaptive Support Weight Approaches

A. Hosni, M. Bleyer, M. Gelautz:
"Near Real-Time Stereo With Adaptive Support Weight Approaches";
Vortrag: International Symposium 3D Data Processing, Visualization and Transmission (3DPVT) 2010, Paris, France; 17.05.2010 - 20.05.2010; in:"International Symposium 3D Data Processing, Visualization and Transmission (3DPVT) 2010", (2010), S. 1 - 8.

[ Publication Database ]

Abstract:


Algorithms based on the adaptive support weight strategy currently represent the state-of-the-art in local stereo matching. Unfortunately, their good-quality results come at the price of high computation times: As opposed to standard local algorithms, incremental computation via sliding windows is not applicable for adaptive support weight windows. This paper presents a method for considerably speeding up computation times of these methods. The key idea is to exploit the adaptive support weight windows for generating an explicit over-segmentation of the reference image in a fast way. Having this explicit segmentation, we can take advantage of a modified"segmentation-based"sliding window technique, which makes run time independent of the window size. In particular, we demonstrate our transformation scheme for the geodesic stereo matcher of \cite{Hosni09} that has recently produced excellent results. Our unoptimized GPU-based implementation processes 320x240 pixel images with 26 allowed disparities at 10 frames per second and achieves rank 32 out of 74 methods in the Middlebury online benchmark.