Performance evaluation and analysis of vanishing point detection techniques

Vanishing point detection algorithms based on a Gaussian sphere representation have been employed in a variety of computer vision systems, for extracting 3D line orientations as a first step towards object detection. Typically, these algorithms have been applied to imagery with strong perspective effects and with little noise or texture, resulting in good solutions for line orientations. However, these algorithms can fail if perspective effects are weak, or if texture edges are predominant; they also fail to take advantage of knowledge about the objects to be detected. In this paper, two new techniques for robust vanishing point detection on the Gaussian sphere are presented; primitive-based vanishing point analysis and interpretation plane error modeling. The performance of these methods, along with two other existing methods from the literature, are quantitatively evaluated and compared for the task of building detection in complex aerial imagery.