Authors:
Mauricio Hess-Flores
1
;
Mark A. Duchaineau
2
;
Michael J. Goldman
2
and
Kenneth I. Joy
1
Affiliations:
1
University of California - Davis, United States
;
2
Lawrence Livermore National Laboratory, United States
Keyword(s):
Dense correspondences, Pose estimation, Scene reconstruction, Bundle adjustment, Resolution pyramid, Error analysis.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Computer Vision, Visualization and Computer Graphics
;
Geometry and Modeling
;
Image-Based Modeling
;
Matching Correspondence and Flow
;
Motion, Tracking and Stereo Vision
;
Pattern Recognition
;
Software Engineering
;
Stereo Vision and Structure from Motion
Abstract:
A novel method to detect and correct inaccuracies in a set of unconstrained dense correspondences between two images is presented. Starting with a robust, general-purpose dense correspondence algorithm, an initial pose estimate and dense 3D scene reconstruction are obtained and bundle-adjusted. Reprojection errors are then computed for each correspondence pair, which is used as a metric to distinguish high and low-error correspondences. An affine neighborhood-based coarse-to-fine iterative search algorithm is then applied only on the high-error correspondences to correct their positions. Such an error detection and correction mechanism is novel for unconstrained dense correspondences, for example not obtained through epipolar geometry-based guided matching. Results indicate that correspondences in regions with issues such as occlusions, repetitive patterns and moving objects can be identified and corrected, such that a more accurate set of dense correspondences results from the
feedback-based process, as proven by more accurate pose and structure estimates.
(More)