Light-Dark: A Novel Lightweight Self-supervised Monocular Depth Estimation in the Dark

Self-supervised monocular depth estimation has been widely studied in recent years. In nighttime scenes where the photometric consistency assumption is not met, several solutions have emerged to address this challenge. However, existing monocular depth estimation algorithms for nighttime often require a large-scale model and a significant number of floating-point operations, making it challenging to apply them to practical problems such as autonomous driving. In this paper, we propose a lightweight monocular depth estimation algorithm tailored for nighttime scenes, named Light-Dark. Specifically, we design a lightweight DepthNet incorporating Feature-Fusion blocks and Cross-connections. Additionally, in response to the low-light and high-noise issues in nighttime scenes, we introduce a Noise-Constrained Adaptive Image Enhancement (NCAIE) module. We deploy our model on the edge device Jetson AGX Orin to validate its real-world performance. A series of experiments conducted on nighttime datasets, RobotCar and nuScenes, indicate the effectiveness of our proposed Light-Dark and the equilibrium between lightweight design and accuracy.