DistanceNet: A Deep Learning Study on Point-to-Point Distance Measurement From Single Images

In computer vision image processing, traditional methods that measure distances on planes perpendicular to the camera axis cannot obtain accurate distances between two points from a single image. Therefore, accurately measuring the distance between any two points from a single image, without precise depth information, becomes a challenge for many researchers. This article pioneers the application of deep learning regression for distance measurement by leveraging depth estimation techniques. Firstly, a method for collecting distance datasets based on chessboard corner detection is proposed to provide the training data for DistanceNet. Secondly, the trained DistanceNet using deep learning regression can directly obtain distances between two points from single images, filling the gap in image-based distance measurement applications. Finally, comparative experiments show that the proposed DistanceNet outperforms traditional analytical methods and basic multilayer perceptron (MLP) networks in terms of distance measurement accuracy. On four test datasets (564, 473, 372, and 129), mean absolute error (MAE) values achieved by DistanceNet are 23.07, 28.21, 13.93, and 13.64, respectively.