LLE for SfM: Learning to Enhance the UAV Images for Nighttime Urban 3D Reconstruction

Despite the ubiquity of unmanned aerial vehicles (UAVs) in urban monitoring, nighttime UAV imagery—plagued by high level of noise, uneven light, motion blur, and low dynamic range—poses unique challenges for 3D reconstruction. These degradations cripple traditional Structure from Motion (SfM) frameworks through unreliable feature extraction and inconsistent cross-view spatial association. We propose a 3D reconstruction framework to recover accurate camera poses and point clouds from unordered UAV images captured in nighttime urban scenes. Traditional SfM pipelines typically rely on well-lighted scene imagery, which becomes challenging in low-light conditions due to poor keypoint detection and matching, potentially leading to a breakdown of the entire SfM system. To address these challenges, we introduce a Low-Light Enhancement SfM (LLE-SfM) framework that leverages an enhancement network to suppress noise while restoring scene colors and texture details. Specifically, our framework first decomposes every image of a set of nighttime UAV images into high/low-frequency layers. The High-frequency Stream is employed to remove noise and enhance edge details, while transformer layers are deployed within the Low-frequency Stream to reconstruct scene colors and refine consistency. Due to the lack of low-light 3D reconstruction benchmark datasets for nighttime urban scenes, we introduce the LLUrban3d dataset, including the Rendered subset and Real subset, to provide sufficient ground truth for the training and evaluation of low-light 3D reconstruction pipeline. Experiments on the LLUrban3d dataset demonstrate that our framework significantly improves reconstruction accuracy and completeness over SfM baseline in nighttime urban scenes, with a maximum improvement of 24.17% in F-score.