Improving Monocular Depth Estimation in Traffic Scenarios by Polarization

Depth estimation in traffic scenarios is a significant task of environment perception in autonomous driving. The traditional methods use monocular RGB images to predict depth. However, with the development of sensor technologies, polarization cameras have emerged. Since polarized images contain more information than normal RGB images, it is necessary to explore the depth estimation based on such images. In this work, we firstly establish a new dataset which collects polarized images of four different scenes including both day- and nighttime, while sparse depth values were collected at the same time. In total, our polarization dataset contains over 3k frames of polarized images. Based on experiments on this dataset, we demonstrate that the performance of monocular depth estimation can be improved by nearly 20% with the aid of polarized images, revealing the great potential of polarized images for depth estimation tasks. The code of our method will be made public at https://github.com/MandeISun/PolarDepthVal.