Real-time Vanishing Point Tracking in Manhattan World Using Improved BaySAC

Abstract

Vanishing points provide geometric information about a scene and assist in camera calibration, scene understanding, and 3D reconstruction. The random sample consensus (RANSAC) algorithm faces challenges of low efficiency and insufficient robustness. With prior information, the Bayesian sample consensus (BaySAC) algorithm can efficiently derive the correct parameters and compensate for the deficiencies of the RANSAC algorithm. This study proposes an improved BaySAC vanish - ing point detection algorithm, which uses a linear grouping strategy to enhance the distribution of inliers across groups and accelerate convergence. In the continuous frame vanishing point tracking problem, the detection result of the previous frame is employed as a priori infor - mation for the subsequent frames, facilitating efficient convergence for vanishing point detection and tracking in videos. Experimental results on both benchmark datasets and real-world image sets demonstrate that the proposed method achieves remarkable accuracy and efficiency, enabling real-time performance for vanishing point detection and track- ing. The code is available at https://github. com/CHEMYE/BaySAC.