A New Method for Vanishing Points Detection in 3d Reconstruction from a Single View

From a single view that contains a three dimensional scene is possible to obtain substantial information about the object and even its partial reconstruction. This analysis can provide, in the case of Photogrammetry, measures of the original object, and in the case of Computer Vision, useful information to 3D reconstruction. Nevertheless, to carry out this analysis, a critical part of the process is to identify with a high degree of quality, the structural elements of an oblique image, mainly vanishing lines and points. The presence of mini-segments due to the automatism, and the own weakness of the scene’s geometry, provide unfavorable intersection cases. Taking into account the qualitative and quantitative gap of complexity associated to real images corresponding to architectural buildings, this paper aims to present a new method for vanishing points detection that takes a scientific approach combining several proven techniques and which applies three different steps: 1. A Clustering step, in order to cluster the mini-segments in vanishing lines. 2. An Estimation step, which applies a modification of the Gaussian sphere method, in order to obtain an estimation of vanishing points and reject possible erroneous vanishing lines. 3. A Computation step, which applies a re-weighted least square adjustment support by an M-estimator.