Globally optimal line clustering and vanishing point estimation in Manhattan world

Optimal Geometric Fitting under the Truncated L2-Norm

“…For example, in [9,8], robust estimators for 3D reconstruction is proposed and in [13], a formulation based on mixed integer programming is given. Similar ideas are presented in [3] for line clustering and vanishing point detection. These methods do not depend on initialization and converge to a global optimum.…”

A Global Approach for the Detection of Vanishing Points and Mutually Orthogonal Vanishing Directions

“…In this case, the VP detection is no longer a problem of multiple model fitting, but the problem of fitting a single triplet of mutually orthogonal VPs in the presence of edges that are outliers. Such fitting can be accomplished through EM [6], by using minimal solutions as hypothesis generator in a RANSAC paradigm [17], or by applying Branch-and-Bound to solve a consensus set maximization that assures global optimality [4]. The disadvantages of this type of approach are that additional VDs that might exist are passed undetected, and the methods cannot handle images with more than one set of Manhattan-world directions for which multi-model fitting is again required (see Fig.…”

Edge Based 3d Indoor Corridor Modeling Using a Single Image

“…Vanishing points are valuable for camera calibration (Kosecka and Zhang 2002;Cipolla et al, 1999;Caprile andTorre 1990, andTardif 2009), estimation of rotation angles (Kosecka and Zhang 2002;Antone andTeller 2000, andDenis et al 2008), and more importantly 3D reconstruction (Parodi andPiccioli 1996, andCriminisi et al, 2000). In order to find vanishing points, different methods of straight line clustering are available (Bazin et al, 2012). There are four main categories for these methods based on: 1) Hough Transform (HT), 2) Random Sample Consensus (RANSAC), 3) Exhaustive Search on some of the unknown entities, and 4) Expectation Maximization (Bazin et al, 2012).…”

“…In order to find vanishing points, different methods of straight line clustering are available (Bazin et al, 2012). There are four main categories for these methods based on: 1) Hough Transform (HT), 2) Random Sample Consensus (RANSAC), 3) Exhaustive Search on some of the unknown entities, and 4) Expectation Maximization (Bazin et al, 2012).…”