REVIEW 3 major objections 6 minor 17 references
Survey on Monocular Metric Depth Estimation
T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Monocular metric depth estimation has become a field of its own, and this survey organizes its datasets, models, and open problems into one reference.
desk verdict A readable but unreliable survey: the anchoring tables have citation and classification errors that undercut its own comprehensiveness claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central organizing object is the distinction between relative depth and metric depth, operationalized through scale-invariant versus scale-aware training and evaluation. The survey's load-bearing mechanisms are its comparison tables: Table 1's timeline of MMDE methods, Table 2's zero-shot benchmark numbers reproduced from Depth Pro (Bochkovskii et al., 2024) across eight datasets, and Table 3's 38-dataset taxonomy. These tables let the survey argue that dataset choice and metric-supervision type shape model behavior, and that no standardized protocol yet exists for fair cross-model comparison.
What would settle it
Re-run the eight zero-shot benchmarks in Table 2 with the official checkpoints of Depth Anything, Depth Anything V2, Metric3D, Metric3D v2, PatchFusion, UniDepth, ZeroDepth, ZoeDepth, and Depth Pro using the original evaluation protocols; if the reproduced numbers diverge substantially from Table 2, the survey's comparative evaluation is not trustworthy. As a lighter check, verify whether BDD100K and Mapillary Vistas provide metric ground truth; if they do, Table 3's 'Relative' labels are wrong and the claim that 32 of 38 datasets are metric needs revision.
Extended reading notes
Core claim
On the paper's own terms, the discovery is a synthesis: the field has moved from ordinal, scale-invariant predictions to models that output metric depth, and the decisive ingredients are dataset scale and diversity together with architectural choices such as adaptive depth bins, camera-intrinsic normalization, patch-based fusion, and diffusion-based refinement. The evidence is organized in three tables: a timeline of key MMDE methods, a zero-shot benchmark comparison across eight datasets using numbers reported by Depth Pro (Bochkovskii et al., 2024), and a taxonomy of 38 datasets labeled by scene type, real or synthetic origin, modality, and whether the labels are metric or relative. The survey also establishes that zero-shot generalization is the leading open problem and that no standardized evaluation protocol yet exists for comparing models fairly.
Load-bearing premise
The load-bearing premise is that the benchmark numbers in Table 2 and the dataset labels in Table 3 are accurate as reported, even though the survey did not independently verify them.
Editorial extensions
If this is right
- A practitioner selecting a model can use the survey's comparison to choose between single-inference speed, patch-based detail, and generative fidelity, since the survey shows these are the three current trade-off regimes.
- Dataset choice is load-bearing: with 32 of 38 datasets providing metric depth, supervised MMDE training has enough raw material, but the six relative-depth datasets still matter for scale-agnostic pretraining.
- The survey's timeline implies that the next advance will likely combine large-scale metric supervision with an architecture that removes camera-intrinsic dependence, since that is the direction shared by ZoeDepth, UniDepth, and DAC.
- Because generative models currently produce mostly relative depth, the survey leaves open the possibility that metric diffusion models will be the next frontier once inference cost and release of checkpoints are addressed.
Reading between the lines
- A reader who wants to compare models should treat Table 2 as a transcription of Depth Pro's evaluation rather than an independent benchmark: the survey's own Section 6.4 says the absence of standardized protocols hinders fair comparison, so the numbers are best read as indicative.
- The classification of BDD100K and Mapillary Vistas as relative-depth datasets in Table 3 is worth revisiting, because dataset releases change and that label determines the paper's count of 32 metric datasets.
- A natural next step, not taken by the survey, would be a meta-analysis correlating dataset properties such as synthetic versus real origin, indoor versus outdoor scenes, and LiDAR density with the zero-shot scores in Table 2 to identify which dataset characteristics actually drive transfer.
- The paper's own remark that DMD is closed-source suggests a reproducibility criterion for future surveys: separate released models from paper-only results when summarizing the state of the art.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This survey reviews monocular metric depth estimation (MMDE), covering the task formulation, classical and deep-learning methods, zero-shot relative-depth precursors such as MiDaS, recent metric-depth architectures (ZoeDepth, Metric3D, UniDepth, Depth Anything, Depth Pro, and others), loss and training strategies, sensor-assisted approaches, and a large dataset overview. The paper argues that existing surveys are either outdated or domain-specific, and that it fills a critical gap by providing a comprehensive, structured reference for MMDE, with comparative performance tables and a dataset taxonomy as central evidence.
Significance. If its comparative data were reliable, the survey would be a useful entry point to the MMDE literature: it is broadly scoped, recent, and organized around practically relevant axes such as zero-shot generalization, boundary preservation, patch-based inference, and generative modeling. The paper also deserves credit for explicitly noting when important claims rest on unavailable artifacts, as in the case of DMD, and for naming the absence of standardized benchmarks as a field-level problem. However, the central claim of being a reliable structured reference depends on the accuracy of the two large comparative tables, and both contain attribution and classification errors that currently undermine that claim. These issues are local and correctable, so the contribution can be salvaged with careful verification.
major comments (3)
- [Table 2 (Section 6.4)] The row labeled 'ZeroDepth' is cited as (Bhat et al., 2023), but Section 5 and the reference list identify ZeroDepth with Guizilini et al. (2023), while Bhat et al. (2023) is ZoeDepth, which already appears as a separate row in the same table. This conflation means the numerical values in the ZeroDepth row may belong to a different model, so the cross-model comparison in Table 2 is not trustworthy as printed. The citation and the numbers need to be verified against the original ZeroDepth and ZoeDepth papers.
- [Table 2 caption / Section 6.4] The table caption states that all results are 'reported by DepthPro (Bochkovskii et al., 2024)'. Relying on a single model paper as the sole source for every entry makes the comparison unverifiable and potentially biased, since DepthPro is itself one of the compared methods. The survey should either reproduce results from each model's original source, cite the benchmark leaderboards, or explicitly qualify the table as a secondary re-reporting with all attendant caveats. Without this, the quantitative claims in Section 6.4 cannot be independently checked.
- [Table 3] Several dataset classifications in Table 3 are incorrect or internally inconsistent. A2D2 is listed as 'Indoor Outdoor' and described as 'including both indoor and outdoor scenes', but A2D2 is an outdoor autonomous-driving dataset. Diode is listed with the cell 'Indoor Indoor' and described as indoor-only, although the Diode benchmark contains both indoor and outdoor scenes. Objaverse is labeled as providing relative depth supervision, but it is a repository of 3D object models without image–depth pairs, so it cannot support the depth-estimation roles claimed for it. These errors directly affect the survey's comprehensiveness claim and the reported count of metric-depth datasets, so every row of Table 3 should be checked against the primary dataset documentation.
minor comments (6)
- [Section 2, first paragraph] The mathematical notation is malformed: 'D := (R)H×W' and 'I := (R)H×W ×3' should use superscripts, and 'pixel ii,j' should read 'pixel i_{i,j}'.
- [Section 4] MiDaS is attributed to (Birkl et al., 2023), but the MiDaS method originates with Ranftl et al.; the cited Birkl et al. paper is the MiDaS v3.1 model zoo. Please cite the original MiDaS paper alongside or instead of the model-zoo paper.
- [Table 3, Diode row] The 'Indoor Outdoor' column for Diode contains the duplicated and contradictory entry 'Indoor Indoor'; this appears to be a formatting error that should be corrected to the dataset's actual scene composition.
- [Figure 2] The label 'MiDAS v3.1' should be 'MiDaS v3.1', and the caption would benefit from stating which hardware and settings were used for the timing and memory measurements.
- [Table 1] The 'Output' column uses 'open' and 'close' to describe code release; 'close' should be 'closed' or 'closed-source' for clarity.
- [Reference list, BinsFormer] The reference entry reads 'arxiv 2022. arXiv preprint arXiv:2203.01502'; the duplicated 'arxiv 2022' should be removed.
Circularity Check
No circularity: the survey synthesizes external published results and contains no derivation, fitting, or prediction that reduces to its own inputs.
full rationale
This paper is a literature survey of monocular metric depth estimation. It does not introduce a model, derive equations, fit parameters, or make predictions from its own fitted quantities. The central claim is that the paper provides a comprehensive, structured review; that claim rests on breadth and accuracy of reporting, not on any chain of derivation. Tables 2 and 3 reproduce numbers and classifications from prior work (e.g., DepthPro's reported benchmarks) and the author's own taxonomy, but reproducing external measurements is not circular: the survey's assertions are independently checkable against the cited sources. The reference list contains no prior work by the paper's author that is invoked as a load-bearing premise, so there is no self-citation chain. Possible inaccuracies in Table 2 (e.g., the ZeroDepth attribution) or Table 3 (e.g., A2D2, Diode, Objaverse labels) are factual/correctness concerns, not circularity, because the survey does not define its conclusions in terms of those entries. Under the stated criteria, and proportionately, no circular step can be exhibited, so the appropriate finding is no significant circularity, score 0.
Assumptions & free parameters
assumptions (3)
- domain assumption The benchmark scores in Table 2 accurately reflect the cited papers' reported results.
- domain assumption The dataset characteristics in Table 3 are correctly classified.
- domain assumption No comparable MMDE-specific survey exists, so the paper fills a critical gap.
Cite this review
Pith. "Pith review of Survey on Monocular Metric Depth Estimation." pith.science (2026). https://pith.science/paper/FSPISMIG
@misc{pith2026250111841,
author = {Pith},
title = {Pith review of: Survey on Monocular Metric Depth Estimation},
year = {2026},
howpublished = {\url{https://pith.science/paper/FSPISMIG}},
note = {Machine review of arXiv:2501.11841}
}
read the original abstract
Monocular Depth Estimation (MDE) enables spatial understanding, 3D reconstruction, and autonomous navigation, yet deep learning approaches often predict only relative depth without a consistent metric scale. This limitation reduces reliability in applications such as visual SLAM, precise 3D modeling, and view synthesis. Monocular Metric Depth Estimation (MMDE) overcomes this challenge by producing depth maps with absolute scale, ensuring geometric consistency and enabling deployment without additional calibration. This survey reviews the evolution of MMDE, from geometry-based methods to state-of-the-art deep models, with emphasis on the datasets that drive progress. Key benchmarks, including KITTI, NYU-D, ApolloScape, and TartanAir, are examined in terms of modality, scene type, and application domain. Methodological advances are analyzed, covering domain generalization, boundary preservation, and the integration of synthetic and real data. Techniques such as unsupervised and semi-supervised learning, patch-based inference, architectural innovations, and generative modeling are evaluated for their strengths and limitations. By synthesizing current progress, highlighting the importance of high-quality datasets, and identifying open challenges, this survey provides a structured reference for advancing MMDE and supporting its adoption in real-world computer vision systems.
Figures
Reference graph
Works this paper leans on
-
[4]
Bochkovskii, A., Delaunoy, A., Germain, H., Santos, M., Zhou, Y ., Richter, S. R., and Koltun, V . Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073,
-
[7]
Guo, Y ., Garg, S., Miangoleh, S. M. H., Huang, X., and Ren, L. Depth any camera: Zero-shot metric depth estimation from any camera. arXiv preprint arXiv:2501.02464,
-
[10]
Inserf: text-driven generative object insertion in neural 3d scenes.arXiv preprint arXiv:2401.05335,
Shahbazi, M., Claessens, L., Niemeyer, M., Collins, E., Tonioni, A., Van Gool, L., and Tombari, F. Inserf: text-driven generative object insertion in neural 3d scenes.arXiv preprint arXiv:2401.05335,
-
[11]
Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion
12 Shriram, J., Trevithick, A., Liu, L., and Ramamoorthi, R. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199,
-
[13]
Outdoor Monocular Depth Estimation: A Research Review
Vyas, P., Saxena, C., Badapanda, A., and Goswami, A. Outdoor monocular depth estimation: A research review. arXiv preprint arXiv:2205.01399,
-
[14]
Tartanair: A dataset to push the limits of visual slam
Wang, W., Zhu, D., Wang, X., Hu, Y ., Qiu, Y ., Wang, C., Hu, Y ., Kapoor, A., and Scherer, S. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4909–4916. IEEE,
work page 2020
-
[15]
Monocular depth estimation based on deep learning: A survey
Xiaogang, R., Wenjing, Y ., Jing, H., Peiyuan, G., and Wei, G. Monocular depth estimation based on deep learning: A survey. In 2020 Chinese Automation Congress (CAC), pp. 2436–2440. IEEE,
work page 2020
-
[16]
Depth anything: Unleashing the power of large-scale unlabeled data
Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., and Zhao, H. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10371–10381, 2024a. Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., and Zhao, H. Depth anything v2. arXiv preprint arXiv:2406....
Show all 17 references
-
[17]
arXiv preprint arXiv:2203.01502,
-
[2014]
Deep ordinal regression network for monocular depth estimation
Fu, H., Gong, M., Wang, C., Batmanghelich, K., and Tao, D. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2002–2011,
2002
-
[2015]
Kick back & relax++: Scaling beyond ground-truth depth with slowtv & cribstv
Spencer, J., Russell, C., Hadfield, S., and Bowden, R. Kick back & relax++: Scaling beyond ground-truth depth with slowtv & cribstv. arXiv preprint arXiv:2403.01569, 2024a. Spencer, J., Tosi, F., Poggi, M., Arora, R. S., Russell, C., Hadfield, S., Bowden, R., Zhou, G., Li, Z.,...
-
[2016]
K., Carneiro, G., and Reid, I
Garg, R., Bg, V . K., Carneiro, G., and Reid, I. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pp. 740–756. Springer,
2016
-
[2019]
Midas v3
Birkl, R., Wofk, D., and Müller, M. Midas v3. 1–a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460,
-
[2021]
Saxena, S., Hur, J., Herrmann, C., Sun, D., and Fleet, D. J. Zero-shot metric depth with a field-of-view conditioned diffusion model. arXiv preprint arXiv:2312.13252,
-
[2022]
F., Birkl, R., Wofk, D., Wonka, P., and Müller, M
Bhat, S. F., Birkl, R., Wofk, D., Wonka, P., and Müller, M. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288,
-
[2023]
Monocular depth estimation: A survey
Bhoi, A. Monocular depth estimation: A survey. arXiv preprint arXiv:1901.09402,
1901 arXiv
-
[2024]
F., and Wonka, P
Li, Z., Bhat, S. F., and Wonka, P. Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10016–10025, 2024a. Li, Z., Bhat, S. F., and Wonk...
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.