Survey on Monocular Metric Depth Estimation (2501.11841v3)

Abstract: Monocular Depth Estimation (MDE) is a core task in computer vision that enables spatial understanding, 3D reconstruction, and autonomous navigation. Deep learning methods typically estimate relative depth from a single image, but the lack of metric scale often leads to geometric inconsistencies. This limitation severely impacts applications such as visual SLAM, detailed 3D modeling, and novel view synthesis. Monocular Metric Depth Estimation (MMDE) addresses this issue by producing depth maps with absolute scale, ensuring frame-to-frame consistency and supporting direct deployment without scale calibration. This paper presents a structured survey of depth estimation methods, tracing the evolution from traditional geometry-based approaches to modern deep learning models. Recent progress in MMDE is analyzed, with a focus on two key challenges: poor generalization and blurred object boundaries. To tackle these problems, researchers have explored various strategies, including self-supervised learning with unlabeled data, patch-based training, architectural enhancements, and generative model integration. Each method is discussed in terms of technical contribution, performance improvement, and remaining limitations. The survey consolidates recent findings, identifies unresolved challenges, and outlines future directions for MMDE. By highlighting key advancements and open problems, this paper aims to support the continued development and real-world adoption of metric depth estimation in computer vision.