FoggyDepth: Leveraging Channel Frequency and Non-Local Features for Depth Estimation in Fog

Abstract

With the development of computer vision technology, unsupervised depth estimation from single images has experienced significant advancements under normal weather conditions, demonstrating highly promising results. Nevertheless, its efficacy in estimating depth under less-than-optimal weather conditions, particularly those characterized by fog, continues to pose substantial challenges. In this paper, we propose FoggyDepth that is designed to utilize channel-wise Fourier transform to remedy this limitation. Specifically, to relieve the problem of photometric consistency assumption not holding in foggy scenes within the unsupervised framework, we employ a channel-dimension Fourier transform to obtain channel global statistical information, thereby enhancing the discriminative ability of global representation. Meanwhile, we generate a series of foggy scene samples corresponding to normal training samples and use them for self-supervised training to guide the model to accurately recover depth in foggy conditions. In addition, to further improve the model performance, we utilize a non-local network to capture long-range spatial dependencies in depth estimation. Comprehensive evaluations conducted on the Oxford RobotCar, nuScenes, and Driving Stereo datasets substantiate the precision and reliability of our proposed method. Through a meticulous comparison with existing leading-edge algorithms in depth estimation, our approach demonstrates superior performance, both qualitatively and quantitatively.