Evaluating Monocular Depth Estimation Methods for High-Quality 3D Reconstruction of Road Environments

Sammanfattning

High-quality 3D reconstruction of road environments is crucial for various applications, including road maintenance, autonomous driving, and urban planning. Monocular Depth Estimation, predicting a 2D-to-3D transformation from a single RGB image, offers a promising, cost-effective alternative to traditional 3D reconstruction methods. This thesis investigates state-of-the-art monocular depth estimation models, focusing on their applicability to 3D road scene reconstruction, specifically utilizing structure-from-motion to address scale ambiguity. Several monocular depth estimation models are evaluated using quantitative and qualitative criteria, including depth accuracy, reconstruction quality, and computational efficiency. Performance is assessed on both real-world image data and synthetic, virtual road scenes to complement the limitations of each dataset. The ground truth for the real-world images are obtained through dense structure-from-motion 3D reconstruction. To ensure a fair and reliable comparison across different models and datasets, the study incorporates pre-processing steps such as masking out irrelevant depth values, ultimately producing more comparable results for the quantitative metrics. The research addresses the challenges of scaling monocular depth estimation outputs with sparse point clouds, evaluating the impact of GPS accuracy on metric measurements and the efficacy of each model in capturing fine details of road surfaces. Results demonstrate that structure-from-motion for accurate rescaling is possible using GPS data and is compared to ideal known camera positions. The most accurate monocular depth estimation results were obtained using the Depth Pro model, closely followed by the Depth Anything V2 model which yielded the best inference time results. Finally, the thesis offers insights into the strengths and limitations of each model, along with directions for future work, particularly regarding the challenges in accurate 3D road surface reconstruction.