Automated Planar Rectification from Repeated Patterns

Repetitive patterns typically arise from man-made objects and are ubiquitous in image collections. Because of their highly repetitive structure, imaged patterns violate statistical assumptions typically made in scene understanding algorithms, often negatively impacting algorithm efficacy. Conversely, repetitive patterns could be detected and modeled so that their distinctive structures are leveraged to uniquely characterize the scenes in which they are present. With that goal in mind, this thesis presents a novel method for the automated detection and sparse 3-D reconstruction of imaged coplanar repetitive patterns. The proposed method applies to a very general class of patterns that encompasses nearly all man-made patterns. Particular contributions include a new set of geometric constraints to eliminate the geometric ambiguity between the imaged and scene pattern, a method to reconstruct the pattern’s motif, and a robust framework that successfully detects and reconstructs patterns in the presence of clutter and imaged from lens distorted cameras.