Regression convolutional network for vanishing point detection

This paper presents a detection method for estimation of vanishing point position with designed regression convolutional neural network. Due to the deep structures of convolutional networks, global high-level features are extracted from the whole image, which helps to locate the vanishing point. In this paper, we provide a new structure of regression neural network based on AlexNet. The structure consists of five convolutional layers, four fully connected layers, Tanh activation function and regression loss function. We feed the neural net with a small number of training dataset and the result proves that this method is adaptable. Compare to classical method, deep learning is more effective on blurred pictures and complex circumstances.