부산대학교 도서관

Self-supervised learning draws much attention to Monocular depth estimation (MDE) since it is free of LiDAR annotations and addresses the daytime domain impressively. However, its performance degrades in challenging environments such as night-time scenes, where the assumptions about uniform lighting are no longer valid. Most methods alleviate this problem by adversarial discriminative learning, e.g., closing the gap between the daytime and night-time domain. This paper will address MDE for the night-time domain utilizing simulation data. We overcome the equivalent camera constraints by an image warping technique, making this approach not require a new benchmark dataset. Since a significant domain shift exists between real-world and synthetic data, we use a novel adversarial learning method to relieve this problem. This work is a pioneer in using synthetic data to solve the MDE problem for night-time scenarios. The experimental results demonstrate that our approach produces a comparable effect to state-of-the-art methods, which proves this approach has potential for future research.