Revisiting Normalized Cross-Correlation for Accurate Camera Pose Estimation and Accurate Real-Time Multiple View Stereo

Almeida, Eduardo de Brito

Full Metadata

Overview

Year:: 2015
Contributor:: Almeida, Eduardo de Brito (creator); Cooper, David (Director); Mundy, Joseph (Reader); Felzenszwalb, Pedro (Reader); Brown University. ENGINEERING: Electrical Sciences and Computer Engineering (sponsor)
Genre:: theses
Subject:: image-based modeling; normalized cross-correlation; multiple view stereo; feature matching; duplicated structures; parallel programming; GPU computing; Computer vision; Parallel programming (Computer science)
Extent:: 21, 250 p.
DOI: https://doi.org/10.7301/Z0TB1580

Files

Description

Abstract:: Normalized cross-correlation (NCC) is a powerful matching tool used for finding corresponding features between two signals. In computer vision applications, NCC can be used to identify corresponding 2-d features among images. In general, a square template is specified around some interest point in an image, and a search for a match is carried out by computing the correlation coefficient between the template pixels and the pixels of similarly sized templates centered at pixels in the second image. Hence, spatial similarity is established and a best match is thus determined. Correlation is expensive to compute on large search areas, especially the normalized version, which is invariant to linear illumination changes. This invariance property is very important in many applications, but the prohibitive cost of computing NCC was not suitable for real-time tasks until the introduction of general purpose graphics processing units (GPUs). This thesis revisits NCC for feature matching applications, including those dominated by other competing methods, and shows that NCC can achieve comparable results or better in terms of accuracy and real-time performance. This thesis focuses on an image-based system for real-time 3-d scene analysis capable of handling general and complex scenes. Three applications that require vast amounts of computation are studied, namely camera pose estimation, match disambiguation for dense repeating features and multiple view stereo (MVS). The principal contributions are: (1) a GPU implementation of NCC that is shared among all applications; (2) an NCC matching method that finds correspondences for camera estimation with better localization accuracy than does the widely popular SIFT matching; (3) a novel matching method for dense repeating features commonly found in architectural scenes that improves camera estimation accuracy even further; and (4) a robust parallel MVS algorithm using NCC that runs in real-time entirely on a GPU, handles aerial scenes and is comparable in accuracy to the slower state-of-the-art using a single CPU. Experiments and evaluations are provided for all of these tasks. NCC is neither rotation- nor scale-invariant, which is dealt with by assuming smooth inter-frame camera motion.
Notes:: Thesis (Ph.D. -- Brown University (2015)

Content

Access Conditions

Rights: In Copyright
Restrictions on Use: Collection is open for research.

Citation

Almeida, Eduardo de Brito, "Revisiting Normalized Cross-Correlation for Accurate Camera Pose Estimation and Accurate Real-Time Multiple View Stereo" (2015). Electrical Sciences and Computer Engineering Theses and Dissertations. Brown Digital Repository. Brown University Library. https://doi.org/10.7301/Z0TB1580

Relations

Collection:

Electrical Sciences and Computer Engineering Theses and Dissertations

Theses and Dissertations for the Electrical Sciences and Computer Engineering department.
...