Self-supervised monocular depth estimation on construction sites in low-light conditions and dynamic scenes

Abstract

Estimating construction scene depth from a single image is crucial for various downstream tasks. Self-supervised monocular depth estimation methods have recently achieved impressive results and demonstrated state-of-the-art performance. However, the low-light conditions and dynamic scenes on construction sites pose significant challenges to these methods, hindering their practical deployment. Therefore, an architecture called LLD-Depth is presented to address these challenges, including an improved ForkGAN model to generate paired low-light images from clear-day images, a new unifying learning method for accurately estimating monocular depth, motion flow, camera ego-motion, and its intrinsic parameters, as well as a training framework to estimate monocular depth under both low-light and clear-day conditions effectively. Finally, the effectiveness of monocular depth estimation in construction scenes is verified. LLD-Depth brings 16.67% and 20.17% gain in relative mean error for clear-day and low-light scenes and 2.60% and 1.80% gain in average order accuracy, achieving state-of-the-art performance.