Preview

PDF, English - main document Download (7MB) | Lizenz: Lighting Estimation in Outdoor Scenes by Lee, Haebom underlies the terms of Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 Germany

Abstract

This dissertation, titled “Lighting Estimation in Outdoor Scenes”, explores the vital aspect of light in computer vision, with a focus on the dynamic and complex nature of outdoor lighting. The work is motivated by the challenges in accurately capturing and interpreting outdoor lighting conditions, which are critical for applications in augmented reality (AR), 3D reconstruction, and autonomous driving technologies. Traditional single-image lighting estimation approaches often fall short due to issues like noise, inconsistency, and the complex interplay of natural elements. This thesis proposes new methodologies that extend beyond these limitations by incorporating both spatial and temporal analyses of lighting. This holistic approach allows for a more accurate and realistic interpretation of outdoor scenes, aiming to improve the realism of virtual objects in AR and the accuracy of various computer vision tasks. The dissertation makes two major contributions: First, it explores the combination of intrinsic image decomposition and lighting estimation through a U-Net architecture, aiming to dissect images into albedo and shading components. While this exploration did not yield publication-worthy results, it provided valuable insights for future research. Second, it introduces advanced spatio-temporal outdoor lighting estimation methodologies, including a four-stage method and an end-to-end model utilizing a Transformer architecture for robust global sun direction estimation. These contributions signify an advancement in lighting estimation, with implications for various real-world applications.