Stereo matching through squeeze deep neural networks

Abstract

Visual depth recognition through Stereo Matching is an active field of research due to the numerous applications in robotics, autonomous driving, user interfaces, etc. Multiple techniques have been developed in the last two decades to achieve accurate disparity maps in short time. With the arrival of Deep Leaning architectures, different fields of Artificial Vision, but mainly on image recognition, have achieved a great progress due to their easier training capabilities and reduction of parameters. This type of networks brought the attention of the Stereo Matching researchers who successfully applied the same concept to generate disparity maps. Even though multiple approaches have been taken towards the minimization of the execution time and errors in the results, most of the time the number of parameters of the networks is neither taken into consideration nor optimized. Inspired on the Squeeze-Nets developed for image recognition, we developed a Stereo Matching Squeeze neural network architecture capable of providing disparity maps with a highly reduced network size without a significant impact on quality and execution time compared with state of the art architectures. In addition, with the purpose of improving the quality of the solution and get solutions closer to real time, an extra refinement module is proposed and several tests are performed using different input size reductions.