论文信息 - Optimization and Performance Evaluation of Stereo-Matching Software on Many-core Processors

Optimization and Performance Evaluation of Stereo-Matching Software on Many-core Processors

Abstract The high-performance Geospatial Information System (GIS) is expected to provide an innovative infrastructure for Earth sciences, enabling near-real-time data and processing services. Stereo-matching software has often been used in generating a Digital Elevation Model (DEM) from a pair of satellite imagery data sets to compute height from a parallax views using two photographic images. There is a need to reduce the computation time required for processing large images. We optimize stereo-matching software on multi-core/many-core processors, including Xeon, Cell and GPGPU. We describe optimization approaches of the correlation calculation part which occupied about 55% of the ovarall computation time. After porting and optimizing software for multi-core/many-core processors, we achieved processing time of 4.79 second (Xeon), 2.28 second (Cell) and 0.97 second (GPGPU) on each platform.

[1] Luigi di Stefano,et al. A fast area-based stereo matching algorithm , 2004, Image Vis. Comput..

[2] Ramin Zabih,et al. Non-parametric Local Transforms for Computing Visual Correspondence , 1994, ECCV.

[3] Laurent Moll,et al. Real time correlation-based stereo: algorithm, implementations and applications , 1993 .

[4] Yasushi Yamaguchi,et al. Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) , 1998, IEEE Trans. Geosci. Remote. Sens..

[5] Rob van Nieuwpoort,et al. Using many-core hardware to correlate radio astronomy signals , 2009, ICS.

[6] Isao Kojima,et al. Design Principles and IT Overviews of the GEO Grid , 2008, IEEE Systems Journal.

[7] Dong-Chul Park,et al. Rooftop Detection Based on 3D Line Data Using Fast Graph Search , 2009, 2009 Ninth International Conference on Hybrid Intelligent Systems.