A critical synthesis of uncertainty quantification and foundation models in monocular depth estimation

Recent foundation models have led to major breakthroughs in monocular depth estimation (MDE), enabling dense 3D reconstruction from single-view imagery across a wide range of domains. However, safe and reliable deployment — particularly for metric depth estimation, which involves predicting absolute distances in real-world units — remains an open challenge. Even state-of-the-art foundation models are prone to critical errors, raising concerns for applications in photogrammetry, robotics, and remote sensing. Equipping these massive Vision Transformers with uncertainty quantification (UQ) is highly non-trivial; standard UQ methods often fail to scale, and it remains an open question how internet-scale pre-training and subsequent metric fine-tuning alter a model’s uncertainty quality. To address these limitations, we integrate five scalable UQ methods with the DepthAnythingV2 foundation model to obtain pixel-wise uncertainty estimates alongside metric depth predictions. We evaluate these combinations across four diverse datasets, including indoor scenes, outdoor urban environments, synthetic-to-real data, and high-resolution aerial imagery. Our findings identify fine-tuning with the Gaussian Negative Log-Likelihood Loss as particularly promising, yielding reliable uncertainty estimates while preserving predictive performance and computational efficiency. By successfully adapting UQ to foundation models, this work advances toward more interpretable and trustworthy MDE pipelines, with direct implications for real-world photogrammetric workflows relying on single-view imagery.