PADENet: An Efficient and Robust Panoramic Monocular Depth Estimation Network for Outdoor Scenes

Keyang Zhou, Kaiwei Wang, Kailun Yang

Abstract: Depth estimation is a basic problem in computer vision, which provides three-dimensional information by assigning depth values to pixels. With the development of deep learning, researchers have focused on estimating depth based on a single image, which is known as the “monocular depth estimation” problem. Moreover, panoramic images have been introduced to obtain a greater view angle recently, but the corresponding model for monocular depth estimation is scarce in the state of the art. In this paper, we propose PADENet for panoramic monocular depth estimation and re-design the loss function adapted for panoramic images. We also perform model transferring to panoramic scenes after training. A series of experiments show that our PADENet and loss function can effectively improve the accuracy of panoramic depth prediction while maintaining a high level of robustness and reaching the state of the art on the CARLA Dataset.