Survey on Monocular Metric Depth Estimation

Jiuling Zhang, Yurong Wu, Huilong JiangComput 2025 · DOI

Abstract

Monocular Depth Estimation (MDE) has evolved from predicting relative depth orders to estimating absolute metric distances from a single RGB image. This transition, known as Monocular Metric Depth Estimation (MMDE), is critical for applications in autonomous navigation, robotics, and 3D reconstruction. This paper surveys the recent shift from geometry-based methods to deep learning foundation models, analyzing how architectural innovations, large-scale datasets, and generative modeling are overcoming the inherent scale ambiguity of 2D projections.

Introduction: The Challenge of Scale

Monocular Depth Estimation is fundamentally an "ill-posed" problem. Because a 2D image is a projection of a 3D scene, the absolute geometric scale is lost during capture. While traditional deep learning models became proficient at relative depth estimation—identifying that one object is behind another—they historically struggled with metric accuracy. For high-stakes applications like visual SLAM, autonomous driving, or surgical robotics, knowing a chair is "farther away" is insufficient; the system must know it is exactly 2.5 meters away.

This survey explores the methodologies bridging the gap between "looks right" (relative) and "is right" (metric), focusing on how the field is moving toward zero-shot generalization across diverse environments.

Methodological Pillars of MMDE

Architectural Innovations

The transition from standard Convolutional Neural Networks (CNNs) to Vision Transformers (ViTs) has been a turning point, allowing models to capture the global context necessary for scale inference. A key technique in this space is Adaptive Binning (e.g., AdaBins, ZoeDepth), which dynamically divides the depth range into discrete, scene-dependent "bins" to increase prediction precision.

Zero-Shot Generalization and Foundation Models

Modern MMDE relies on "foundation models" trained on massive, heterogeneous datasets. These models aim for zero-shot generalization, meaning they can accurately predict depth in unseen domains (e.g., an indoor model working on an outdoor highway) without requiring per-scene calibration.

High-Resolution and Patch-Based Inference

To maintain fine structural details in high-resolution images, methods such as PatchFusion and DepthPro utilize patch-based processing. By dividing images into overlapping tiles and fusing the results, these models avoid the blurring associated with downsampling, though they often face increased computational latency.

Generative Modeling and Diffusion

A significant emerging trend is the use of Diffusion Models (e.g., DMD, Marigold). Unlike discriminative models that predict depth in a single forward pass, diffusion models iteratively refine depth maps. This approach often yields significantly sharper boundaries and better handles challenging surfaces, such as transparent or reflective materials, which typically confuse standard sensors.

Geometric Grounding

Newer frameworks like Metric3D and UniDepth explicitly incorporate camera intrinsics (such as focal length). By mapping images into a "canonical space," these models ensure that the output scale is physically grounded and consistent with the camera's optical properties.

The Data Revolution

Progress in MMDE is driven as much by data as by architecture. The survey categorizes the data landscape into four primary types:

  1. Outdoor/Driving: Datasets like KITTI and Waymo Open Dataset provide LiDAR-based ground truth for high-speed environments.
  2. Indoor: NYU Depth V2 and ScanNet utilize Time-of-Flight (ToF) or Kinect sensors for close-quarters mapping.
  3. Synthetic: TartanAir and Hypersim provide "pixel-perfect" labels. High-quality synthetic data has become essential for bridging the "domain gap," as it allows models to learn sharp object boundaries without sensor noise.
  4. Large-Scale Semi-Supervised: Models like Depth Anything leverage massive unlabeled datasets (up to 62 million images) to achieve unprecedented robustness.

Key Insights and Performance Benchmarks

The Relative-to-Metric Bridge

A major breakthrough was the development of ZoeDepth, which demonstrated that a pre-trained relative depth model (like MiDaS) could be effectively fine-tuned into a metric model using a lightweight head and scene-aware routing. This proved that relative features contain significant latent metric information.

Current State-of-the-Art

As of recent benchmarks, Metric3D v2 and Depth Anything V2 lead the field in zero-shot performance. Metric3D v2 shows exceptional accuracy in outdoor settings (achieving a low AbsRel of 3.9 on KITTI), while Depth Anything V2 excels in indoor generalization. However, a clear tradeoff remains: while single-pass models are fast enough for real-time robotics, generative and patch-based models—which produce the highest fidelity results—can be 10 to 100 times slower.

Limitations and Future Frontiers

Despite rapid progress, several challenges remain:

  • Temporal Consistency: Most MMDE research focuses on static images. In video sequences, this often results in "flickering" depth, which is problematic for stable navigation.
  • Edge Artifacts: Even top-tier models struggle with "halo" effects around thin objects or complex occlusions.
  • Computational Efficiency: Bringing the high-fidelity results of diffusion models to real-time speeds is a primary goal for the next generation of researchers.
  • Standardization: The field lacks a unified training protocol, making it difficult to compare models that use different data augmentations and loss functions.

Conclusion

Monocular Metric Depth Estimation has moved from a theoretical challenge to a practical reality. By combining the global context of Transformers, the refinement of Diffusion models, and the scale of massive datasets, the field is successfully resolving the scale ambiguity problem, paving the way for more capable and aware autonomous systems.

Summary updated

BibTeX
@article{zhang2025survey,
  author = {Jiuling Zhang and Yurong Wu and Huilong Jiang},
  title = {{Survey on Monocular Metric Depth Estimation}},
  journal = {Comput.},
  year = {2025},
  doi = {10.3390/computers14110502},
}
Figure