Shadow vanishing point detection via combined human/shadow adaptive modulation

Abstract

A set of parallel lines in 3D space intersect at a common vanishing point when projected to a 2D image. Vanishing point detection plays an important role in 3D computer vision and previous works are usually focused on the parallel lines in man-made structures, such as buildings, railways, and lanes. Under sunlight, shadow directions in many instances, especially standing persons, are largely parallel on the ground. The detection of human/shadow vanishing points in 2D images can well complement others for 3D information, especially in scenes that lack man-made structures. However, this is a very challenging problem since human shadows do not present explicit parallel line features along the shadow directions. In this paper, we propose a new ShadowVPNet for shadow vanishing point detection, which consists of a combined human/shadow detector, a feature extractor and a vanishing point classifier. Adaptive modulation modules conditioned on the combined human/shadow priors are incorporated into the feature extractor for modulating the extracted features. We construct a new shadow vanishing point image dataset, as well as all the needed ground-truth annotations, for network training and testing. Experimental results demonstrate the effectiveness of the proposed method on shadow vanishing point detection. The code and dataset are released at https://github.com/hhqweasd/SVPD.