Self-Supervised Cross-Connected CNNs for Binocular Disparity Estimation

Abstract

When developing autonomous vehicles, sensors with high accuracy and speed are needed. One type of sensor that can gather a lot of information is the camera. From two stereo images a disparity between them can be calculated, and from that the depth. The drawback with today’s algorithms is the trade-off between high quality estimation and computational speed. By taking inspiration from recently published neural networks for other applications, we present a novel design for disparity estimation networks. We design a cross-connected convolutional neural network to calculate full HD disparity maps from stereo images at a high frequency. By transfer training the network, using self-supervised learning, the network can learn to handle new environments. The network shows significantly faster runtimes than other disparity estimation networks, with the loss of some accuracy. We show that the self-supervised loss functions perform poorly when the images are not aligned, which is important to solve for real life applications of the network. Furthermore, we present ideas on how to improve the network’s runtime further.