Spatio-temporal Super-resolution with photographic and depth data using GANs

Abstract

LiDAR technology is essential for self-driving cars, which have seen a surge in interest and investments from startups and established automotive corporations alike. However, the task of automated driving requires high resolution and significant depth-range capabilities of the sensor, keeping its cost prohibitive. Super-resolution of depth maps has been explored as a potential circumvention of these problems, with a substantial number of methods being analyzed in the past few years, yielding various levels of success. We propose a super-resolution algorithm trained for depth-map data and LiDAR compatibility using Generative Adversarial Networks (GANs).