An End to End Learning Approach for Distance Estimation Trained with Artificially Generated Stereo Images

This paper proposes a solution for distance estimation using stereo images. The solution is a convolutional neural network that takes two images as an input, and outputs the distance estimate, without the need for prior camera calibration or disparity map calculation. The dataset used for training consists of images generated from an artificially constructed 3D scene. The training algorithm used was stochastic gradient descent. Evaluation of the solution was conducted on a separate dataset. Mean absolute error after the evaluation was 1.59 m, while the median value of the absolute error was 1.2 m. These results show that the proposed solution is a valid proof of concept for the usage of convolutional neural networks for the distance estimation of objects in stereo images in a single step.