Efficient Techniques for High Resolution Stereo Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 20, 2019
  • Wang, Yilin
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • The purpose of stereo is extracting 3-dimensional (3D) information from 2-dimensional (2D) images, which is a fundamental problem in computer vision. In general, given a known imaging geometry the position of any 3D point observed by two or more different views can be recovered by triangulation, so 3D reconstruction task relies on figuring out the pixel’s correspondence between the reference and matching images. In general computational complexity of stereo algorithms is proportional to the image resolution (the total number of pixels) and the search space (the number of depth candidates). Hence, high resolution stereo tasks are not tractable for many existing stereo algorithms whose computational costs (including the processing time and the storage space) increase drastically with higher image resolution. The aim of this dissertation is to explore techniques aimed at improving the efficiency of high resolution stereo without any accuracy loss. The efficiency of stereo is the first focus of this dissertation. We utilize the implicit smoothness property of the local image patches and propose a general framework to reduce the search space of stereo. The accumulated matching costs (measured by the pixel similarity) are investigated to estimate the representative depths of the local patch. Then, a statistical analysis model for the search space reduction based on sequential probability ratio test is provided, and an optimal sampling scheme is proposed to find a complete and compact candidate depth set according to the structure of local regions. By integrating our optimal sampling schemes as a pre-processing stage, the performance of most existing stereo algorithms can be significantly improved. The accuracy of stereo algorithms is the second focus. We present a plane-based approach for the local geometry estimation combining with a parallel structure propagation algorithm, which outperforms most state-of-the-art stereo algorithms. To obtain precise local structures, we also address the problem of utilizing surface normals, and provide a framework to integrate color and normal information for high quality scene reconstruction.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Frahm, Jan-Michael
  • Dunn, Enrique
  • Fuchs, Henry
  • Mordohai, Philippos
  • Niethammer, Marc
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2016

This work has no parents.