Dominant vanishing point detection in the wild with application in composition analysis

The vanishing point provides a strong ability to infer the 3D structure of the scene. It finds great application in image composition analysis, lane detection, camera calibration and salience detection. Many methods have been proposed to predict the location of vanishing point. They are usually based on geometrical and structural features such as lines or contours. However, such methods suffer deteriorated accuracy due to the large number of outlier line segments in natural landscape images. In this paper, we propose a semantic-texture fusion network to detect the dominant vanishing point in the image. The proposed network includes two branches. The first branch is based on the Holistically-Nested Edge Detection Network which extracts textural features. The second branch aims to extract the semantic features. In order to boost the representational power of a network, we adopt the Squeeze-and-Excitation block to model the interdependencies between the semantic features and the textural features. Experimental results reveal a step forward against the state-of-the-art vanishing point detection methods in natural landscapes. Based on the detection results, we further demonstrate how the proposed model can be used to provide on-line guidance to amateur photographers.