Single-Image Depth Estimation in Low-Light or Night-Time Conditions

Abstract

Depth estimation is a fundamental component in a variety of computer vision tasks. However, models trained solely on daytime imagery often underperform when applied to night-time scenes due to significant domain shifts and altered visual characteristics. To address this issue, we generated a synthetic night-time dataset using image translation techniques powered by a generative neural network. This synthetic data was then employed to fine-tune a pre-existing depth estimation model. Our goal was to assess whether training with generated night-time images could enhance the model’s accuracy in such conditions. Through comparative evaluations, we found that the fine-tuned model delivered substantially better performance on night-time images, showing results on par with methods explicitly built for daytime scenarios. These outcomes emphasize the value of synthetic training data in bridging the performance gap for depth estimation under low-light conditions.