DNA-Depth: A Frequency-Based Day-Night Adaptation for Monocular Depth Estimation

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; pp. 1 - 12
• Main Authors: Shen, Mengjiao, Wang, Zhongyi, Su, Shuai, Liu, Chengju, Chen, Qijun
• Format: Journal Article
• Language: English
• Published: New York IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Algorithms; Cameras; Daytime; Depth estimation; domain adaptation; dynamic environment; Estimation; Fast Fourier transformations; Fourier transform; Fourier transforms; Frequency estimation; Frequency-domain analysis; Light sources; Lighting; monocular vision; Motion simulation; Night; Optical flow; Optical flow (image analysis); Optimization; Training
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2023.3322498

Autonomous driving necessitates ensuring safety across diverse environments, particularly in challenging conditions like low-light or nighttime scenarios. As a fundamental task in autonomous driving, monocular depth estimation has garnered significant attention and discussion. However, current monocular depth estimation methods primarily rely on daytime images, which limits their applicability to nighttime scenarios due to the substantial domain shift between daytime and nighttime styles. In this article, we propose a novel Day-Night Adaptation method (DNA-Depth) to realize monocular depth estimation in a night environment. Specifically, we simply use Fourier Transform to address the domain alignment problem. Our method does not require extra adversarial optimization but is quite effective. The simplicity of our method makes it easy to guide day-to-night domains. To the best of our knowledge, we are the first to utilize fast Fourier transformation for nighttime monocular depth estimation. Furthermore, to alleviate the problem of mobile light sources, we utilize an unsupervised joint learning framework for depth, optical flow, and ego-motion estimation in an end-to-end manner, which is coupled by 3-D geometry cues. Our model can simultaneously reason about the camera motion, the depth of a static background, and the optical flow of moving objects. Extensive experiments on the Oxford RobotCar, nuScenes, and Synthia datasets demonstrate the accuracy and precision of our method by comparing it with those state-of-the-art algorithms in depth estimation, both qualitatively and quantitatively.