Pith. sign in

REVIEW 4 major objections 6 minor 81 references

Underwater Monocular Metric Depth Estimation: Real-World Benchmarks and Synthetic Fine-Tuning with Vision Foundation Models

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fine-tuning a depth foundation model on physics-simulated underwater images improves metric accuracy on real underwater benchmarks.

desk verdict A genuinely useful underwater depth benchmark whose headline fine-tuning claim needs a same-recipe control before it can be trusted. read the letter →

arxiv 2507.02148 v2 pith:V2TAIUSZ submitted 2025-07-02 cs.CV

Zijie Cai , Christopher Metzler This is my paper
classification cs.CV
keywords underwaterdepthestimationmonocularmetricdomainadaptationsyntheticdataphysics-basedimageformationvisionfoundationmodelszero-shotbenchmarkingfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that monocular metric depth models trained on terrestrial data, even large vision foundation models, lose accuracy when applied underwater, and that a targeted synthetic fine-tuning step can recover much of the loss. The fix is to fine-tune Depth Anything V2 (ViT-S) on a synthetic underwater version of the Hypersim indoor dataset, generated by a physics-based image formation model with wavelength-dependent attenuation and backscatter across several Jerlov water types. On the real FLSea and SQUID benchmarks, the fine-tuned model improves $\delta_1$ accuracy over the clean in-air Hypersim baseline on all three test groups, with the largest gains on the more turbid SQUID set. If this holds, low-cost synthetic data can substitute for scarce real underwater metric ground truth in domain adaptation.

What carries the argument

The load-bearing object is the synthetic underwater image formation model: for each channel $c$, the observed intensity is $I_c = J_c \cdot e^{-\beta_c z} + B^\infty_c \cdot (1 - e^{-\beta_c z})$, where $J_c$ is the clear scene radiance, $\beta_c$ the wideband attenuation coefficient, $B^\infty_c$ the veiling light at infinity, and $z$ the per-pixel depth from Hypersim. Applying this to indoor Hypersim RGB-D images, with coefficients drawn from Jerlov open-ocean and coastal water classes, yields paired turbid-looking RGB images and unchanged metric depth maps. The fine-tuning recipe then freezes the first half of the ViT-S encoder and updates the remaining encoder layers and the DPT decoder with SiLog loss at a 20-meter depth scale, with color augmentations, so the model keeps its terrestrial scene priors while learning underwater-specific appearance cues.

What would settle it

Train the same fine-tuning recipe on a synthetic underwater dataset whose attenuation and backscatter coefficients are drawn from a different distribution, or whose scenes are outdoor rather than indoor, and test on FLSea and SQUID: if the $\delta_1$ improvement disappears or reverses, the gain is tied to Hypersim's geometry or to the particular parameter choice rather than to the physics-based underwater image formation model.

Watch

Extended reading notes

Core claim

The paper's central claim is that forward domain adaptation with physics-simulated underwater images transfers to real underwater scenes: fine-tuning Depth Anything V2 (ViT-S) on underwater-rendered Hypersim, rather than on clean in-air Hypersim, reduces AbsRel and raises $\delta_1$ on real-world benchmarks, improving $\delta_1$ on all three groups (FLSea-Canyon, FLSea-Red Sea, SQUID) and AbsRel on two of the three. The mechanism is the wideband underwater image formation model $I_c = J_c e^{-\beta_c z} + B^\infty_c (1 - e^{-\beta_c z})$, applied per RGB channel with Jerlov water-type coefficients, which produces paired underwater RGB images and clean metric depth while preserving pixel alignment. The paper also benchmarks six zero-shot models and finds that while UniDepth V2 (ViT-L) leads on clear-water FLSea, performance drops sharply in turbid SQUID, and that the underwater-specific UW-Depth model outperforms much larger general-purpose models on SQUID despite a compact backbone.

Load-bearing premise

The load-bearing premise is that the simplified wideband underwater image formation model, applied to indoor Hypersim scenes, produces degradations similar enough to real underwater light transport that a model fine-tuned on those synthetic images transfers to real FLSea and SQUID photos; the paper does not report the exact $\beta_c$ and $B^\infty_c$ values used and does not quantitatively measure this synthetic-to-real gap.

Editorial extensions

If this is right

  • Fine-tuning on synthetic underwater Hypersim improves $\delta_1$ accuracy on all three real benchmark groups: FLSea-Canyon from 0.4463 to 0.4683, FLSea-Red Sea from 0.4722 to 0.6170, and SQUID from 0.2054 to 0.3204.
  • The same physics-based degradation can be layered onto any in-air RGB-D dataset, so synthetic underwater supervision is not limited to Hypersim's indoor geometry.
  • The synthetic-to-real gain is largest on SQUID, the more turbid and longer-range benchmark, indicating the approach specifically helps where veiling light and attenuation dominate.
  • A compact underwater-specific model trained on real data outperforms much larger zero-shot foundation models on SQUID, showing that domain-specific training can matter more than model scale in degraded conditions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the transfer holds, the same degradation recipe could be applied to other metric depth foundation models, such as UniDepth V2, or to other dense prediction tasks like surface normal estimation, because the image formation model is task-agnostic.
  • The reported gain pattern suggests the fine-tuning mainly teaches the model to discount backscatter and attenuation rather than to improve geometric reasoning; a testable prediction is that improvements concentrate in high-backscatter, low-contrast regions.
  • Because the exact $\beta_c$ and $B^\infty_c$ values per Jerlov class are not reported, the recipe is not fully reproducible; publishing these coefficients and validating on tank imagery with measured water properties would isolate whether the gain comes from the physics model or from generic color augmentation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper studies monocular metric depth estimation for underwater images. The authors benchmark six models (Depth Anything V2, Metric3D V2, UniDepth V2, ZoeDepth, Depth Pro, and UW-Depth) in a zero-shot setting on the FLSea and SQUID real-world underwater datasets, construct a synthetic underwater dataset by applying the wideband underwater image formation model of Eq. (1) to Hypersim, and fine-tune Depth Anything V2 (ViT-S) on that synthetic dataset. They report that fine-tuning improves delta1 on all three benchmark groups and AbsRel on two of the three, with qualitative improvements shown in Fig. 3. The paper claims consistent improvement across all benchmarks and proposes synthetic underwater fine-tuning as an effective domain-adaptation strategy.

Significance. The paper addresses a practically important problem and makes a useful empirical contribution by quantifying the zero-shot performance of several metric-depth foundation models on real underwater data. The test data are external to the synthetic training set, so the evaluation is not circular. The synthetic data pipeline is simple and potentially scalable, and the qualitative comparisons are informative. However, the central causal claim—that underwater simulation, rather than supervised fine-tuning on Hypersim in general, drives the observed gains—is not yet supported because the baseline is not a same-recipe clean-Hypersim fine-tune. The overstatement in the abstract and the lack of variance estimates further weaken the current presentation. With a proper control and corrected claims, the work would be a solid contribution.

major comments (4)
  1. [§4.4, Table 2; §3.4] The central claim that synthetic underwater fine-tuning causes the observed gains is not identified by the reported comparison. Table 2 compares the authors' fine-tuned model, which is initialized from the official relative-depth checkpoint and trained for 20 epochs with the recipe in §3.4 (AdamW, SiLogLoss, color augmentations, frozen first half of the encoder), against a baseline that is the official Depth Anything V2 Hypersim metric-depth checkpoint described in §3.2, whose training details are not controlled by the authors. Any improvement could therefore come from supervised metric fine-tuning on Hypersim itself or from the augmentations, rather than from the underwater simulation. Please add a same-recipe control trained on clean in-air Hypersim with identical initialization, epochs, augmentations, loss, and depth range, and report its numbers alongside Table 2.
  2. [Abstract; Table 2; §5.2] The abstract and contributions state that fine-tuning 'consistently improves performance across all benchmarks,' but Table 2 shows FLSea-Canyon AbsRel worsening from 0.3576 to 0.3620, and §5.2 itself only claims improvements on 'most FLSea subsets.' The wording should be corrected to match the data, and the Canyon result should be discussed explicitly rather than glossed over. In addition, no error bars, multiple seeds, or significance tests are provided, so it is unclear whether the reported gains (e.g., Red Sea delta1 from 0.4722 to 0.6170) exceed run-to-run variability; please add variance estimates or per-scene scores, especially given the small size of SQUID (57 stereo pairs).
  3. [§3.3, Eq. (1)] The synthetic data generation is not reproducible as described. The paper lists Jerlov water types but never gives the numeric values of beta_c and B_inf_c used in Eq. (1), nor the sampling distribution over water types, nor whether each image is assigned a single water type or a blend. Without these values, readers cannot regenerate the dataset or assess how much of the simulated attenuation range is covered. Please report the full parameter table and sampling procedure, and ideally include a quantitative analysis of the synthetic-to-real domain gap using real underwater attenuation or color statistics from FLSea and SQUID.
  4. [Table 1; §4.3; §5.3] The benchmark coverage is overstated relative to what is reported: Metric3D V2 (ViT-L) appears only in qualitative comparisons because its quantitative results are excluded due to hardware constraints, and UW-Depth is omitted on FLSea because of training-set overlap. The paper should either supply the missing quantitative results for these cases or explicitly restrict the 'comprehensive benchmark' claim to the models and splits for which metrics are actually reported.
minor comments (6)
  1. [§3.4] The sentence 'which each image is one of the Jerlov water classes' should read 'where each image is assigned one of the Jerlov water classes.'
  2. [Figure 1 caption] The caption 'RGB Depth Type I Type II...' is unclear; please separate the panel labels (e.g., 'RGB, Depth, Type I, Type II, ...') so that each image is identified.
  3. [Table 2] The abbreviation 'DA V2' is used in the table but not defined there; please define it at first use in the table or in the surrounding text.
  4. [§5.3] The statement that the fine-tuned Depth Anything V2 model 'ranks just behind UniDepth V2' is not supported by Table 2, where the fine-tuned ViT-S is substantially worse than UniDepth V2 ViT-L on every reported metric; please qualify this as a qualitative observation or compare against a matching backbone size.
  5. [§4.5 and Fig. 3] Figure 3 is described as showing 'much sharper depth boundaries and a more accurate metric depth scale,' but the quantitative gains in Table 2 are modest and inconsistent on FLSea-Canyon; please temper the qualitative language or tie it to a per-scene quantitative analysis.
  6. [General] The paper does not state whether code, the synthetic dataset, or fine-tuned checkpoints will be released; for a domain-adaptation study whose main value is the benchmark and the training recipe, releasing these artifacts would greatly improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons against external real-world benchmarks, not derivations reducible to their inputs.

full rationale

The paper's central contribution is an empirical benchmark plus a domain-adaptation experiment: it fine-tunes Depth Anything V2 (ViT-S) on a synthetic underwater version of Hypersim and measures accuracy on real FLSea and SQUID test images (Section 4.4, Table 2). The evaluation data are external to the training pipeline and are not used to fit any parameter of the synthetic generation model. Equation (1) is a standard wideband underwater image formation model imported from external prior work [1, 10, 27, 56], and the Jerlov water-type coefficients are presented as established optical classifications [58]; no quantity in the model is fitted to the benchmark outputs, so there is no self-definitional or fitted-input-called-prediction structure. No uniqueness theorem, ansatz, or load-bearing self-citation is invoked to force the method choice; the citations to [1] and [71] are background support for the forward model and the pretrained weights, respectively, and neither reduces the measured test improvement to an input by construction. The absence of a same-recipe clean-Hypersim control and the FLSea-Canyon AbsRel regression (0.3576 to 0.3620) are legitimate experimental and reporting concerns about causal attribution, but they are confounds, not circularity: the comparison still evaluates the model against real, external ground truth rather than against the training data or a quantity defined in terms of the prediction. Accordingly, no circular steps are identified and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's contribution is empirical. It relies on the fidelity of the Akkaynak-Treibitz wideband model, the choice of Jerlov water type parameters, the reliability of FLSea and SQUID ground truth, and the representativeness of Hypersim indoor scenes. No new theoretical entities are introduced; the only tuned quantities are the synthetic image formation parameters and standard training hyperparameters.

free parameters (1)
  • Wideband attenuation coefficients beta_c and veiling light B_inf_c per Jerlov water type = not specified in paper; taken from Jerlov water type literature [58]
    These coefficients control the synthetic underwater appearance (Eq. 1) and therefore determine what the fine-tuned model learns; exact values are not listed, so the data-generation recipe is not fully reproducible.
assumptions (4)
  • domain assumption Akkaynak-Treibitz wideband underwater image formation model (Eq. 1) is an adequate forward model for generating training images that transfer to real underwater photography.
    Section 3.3 applies it to Hypersim; the whole fine-tuning claim depends on synthetic-to-real transfer.
  • domain assumption Jerlov water type optical properties cover the range of real FLSea and SQUID water conditions.
    Section 3.3 simulates Jerlov I-III and 1C-9C; no quantitative validation against measured water properties at test sites is provided.
  • domain assumption FLSea SfM and SQUID stereo ground-truth depths are accurate enough to serve as metric benchmarks.
    Section 3.5 describes ground truth via photogrammetry and stereo triangulation; errors in ground truth propagate to all reported metrics.
  • domain assumption Hypersim indoor scene geometry is a sufficient training distribution for underwater depth cues.
    Section 5.5 acknowledges only indoor geometry was used, limiting diversity; this is load-bearing for the fine-tuning result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Underwater Monocular Metric Depth Estimation: Real-World Benchmarks and Synthetic Fine-Tuning with Vision Foundation Models." pith.science (2026). https://pith.science/paper/V2TAIUSZ

@misc{pith2026250702148,
  author       = {Pith},
  title        = {Pith review of: Underwater Monocular Metric Depth Estimation: Real-World Benchmarks and Synthetic Fine-Tuning with Vision Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2TAIUSZ}},
  note         = {Machine review of arXiv:2507.02148}
}
read the original abstract

Monocular depth estimation has recently progressed beyond ordinal depth to provide metric depth predictions. However, its reliability in underwater environments remains limited due to light attenuation and scattering, color distortion, turbidity, and the lack of high-quality metric ground truth data. In this paper, we present a comprehensive benchmark of zero-shot and fine-tuned monocular metric depth estimation models on real-world underwater datasets with metric depth annotations, including FLSea and SQUID. We evaluated a diverse set of state-of-the-art Vision Foundation Models across a range of underwater conditions and depth ranges. Our results show that large-scale models trained on terrestrial data (real or synthetic) are effective in in-air settings, but perform poorly underwater due to significant domain shifts. To address this, we fine-tune Depth Anything V2 with a ViT-S backbone encoder on a synthetic underwater variant of the Hypersim dataset, which we simulated using a physically based underwater image formation model. Our fine-tuned model consistently improves performance across all benchmarks and outperforms baselines trained only on the clean in-air Hypersim dataset. This study presents a detailed evaluation and visualization of monocular metric depth estimation in underwater scenes, emphasizing the importance of domain adaptation and scale-aware supervision for achieving robust and generalizable metric depth predictions using foundation models in challenging environments.

Figures

Figures reproduced from arXiv: 2507.02148 by the authors.

Figure 1
Figure 1. Examples from our synthetic underwater dataset. Top [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparisons on 6 underwater scenes from two real-world underwater datasets: FLSea and SQUID. Each group shows the RGB input, ground-truth (GT), UW-Depth† [18], Metric3D V2† [30], Depth Anything V2 (ViT-L)† [71], Depth Pro [7], UniDepth V2 (ViT-L) [47], and ZoeDepth [6]. UniDepth V2 (ViT-L) [47] consistently produces the most accurate metric depth maps with fine edge details across all datasets. Depth Any… view at source ↗
Figure 3
Figure 3. Qualitative comparison between the baseline (Depth Anything V2 ViT-S [ [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 67 canonical work pages

  1. [1]

    A revised underwater image formation model

    Derya Akkaynak and Tali Treibitz. A revised underwater image formation model. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6723– 6732, 2018. 1, 2, 3, 4, 8

    2018

  2. [2]

    Sea-thru: A method for removing water from underwater images

    Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2

    work page 2019

  3. [3]

    Unav-sim: A visually realistic underwater robotics simulator and synthetic data-generation framework

    Abdelhakim Amer, Olaya ´Alvarez Tu ˜n´on, Halil ˙Ibrahim U˘gurlu, Jonas Le Fevre Sejersen, Yury Brodskiy, and Erdal Kayacan. Unav-sim: A visually realistic underwater robotics simulator and synthetic data-generation framework. In 2023 21st International Conference on Advanced Robotics (ICAR), pages 570–576, 2023. 3

    work page 2023

  4. [4]

    Foundation models defining a new era in vision: A survey and outlook

    Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundation models defining a new era in vision: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 47(4):2245–2264, 2025. 2

    work page 2025

  5. [5]

    Underwater single image color restoration using haze-lines and a new quantitative dataset, 2018

    Dana Berman, Deborah Levy, Shai Avidan, and Tali Treibitz. Underwater single image color restoration using haze-lines and a new quantitative dataset, 2018. 1, 2, 3, 5, 6, 8

    work page 2018

  6. [6]

    Zoedepth: Zero-shot transfer by com- bining relative and metric depth, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot transfer by com- bining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8

    work page 2023

  7. [7]

    Richter, and Vladlen Koltun

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second, 2025. 1, 2, 3, 5, 6, 7

    work page 2025

  8. [8]

    Pyramid stereo matching network

    Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5418,

Show all 81 references
  1. [9]

    When vision transformers outperform resnets without pre- training or strong data augmentations.arXiv preprint arXiv:2106.01548, 2021

    Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre- training or strong data augmentations.arXiv preprint arXiv:2106.01548, 2021. 2

    2021 arXiv

  2. [10]

    Chiang and Ying-Ching Chen

    John Y . Chiang and Ying-Ching Chen. Underwater im- age enhancement by wavelength compensation and dehaz- ing.IEEE Transactions on Image Processing, 21(4):1756– 1769, 2012. 4

    2012

  3. [11]

    Best practices for fine-tuning vi- sual classifiers to new domains

    Brian Chu, Vashisht Madhavan, Oscar Beijbom, Judy Hoff- man, and Trevor Darrell. Best practices for fine-tuning vi- sual classifiers to new domains. InComputer Vision – ECCV 2016 Workshops, pages 435–442, Cham, 2016. Springer In- ternational Publishing. 4

    2016

  4. [12]

    Indoor semantic segmentation using depth in- formation.arXiv preprint arXiv:1301.3572, 2013

    Camille Couprie, Cl ´ement Farabet, Laurent Najman, and Yann LeCun. Indoor semantic segmentation using depth in- formation.arXiv preprint arXiv:1301.3572, 2013. 1, 2, 5

    2013 arXiv

  5. [13]

    Diffusion models in vision: A survey

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 45(9):10850–10869, 2023. 3

    2023

  6. [14]

    Davis, R

    J. Davis, R. Ramamoorthi, and S. Rusinkiewicz. Spacetime stereo: a unifying framework for depth from triangulation. In2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II–359, 2003. 5

    2003

  7. [15]

    CRC Press, 2017

    Pinliang Dong and Qi Chen.LiDAR remote sensing and ap- plications. CRC Press, 2017. 1

    2017

  8. [16]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint ar...

    2010 arXiv

  9. [17]

    Simultaneous local- ization and mapping: part i.IEEE robotics & automation magazine, 13(2):99–110, 2006

    Hugh Durrant-Whyte and Tim Bailey. Simultaneous local- ization and mapping: part i.IEEE robotics & automation magazine, 13(2):99–110, 2006. 1

    2006

  10. [18]

    Metri- cally scaled monocular depth estimation through sparse pri- ors for underwater robots, 2023

    Luca Ebner, Gideon Billings, and Stefan Williams. Metri- cally scaled monocular depth estimation through sparse pri- ors for underwater robots, 2023. 2, 4, 5, 6, 7

    2023

  11. [19]

    Depth map prediction from a single image using a multi-scale deep net- work, 2014

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work, 2014. 2, 4, 5

    2014

  12. [20]

    3-d mapping with an rgb-d camera

    Felix Endres, J ¨urgen Hess, J ¨urgen Sturm, Daniel Cremers, and Wolfram Burgard. 3-d mapping with an rgb-d camera. IEEE transactions on robotics, 30(1):177–187, 2013. 1

    2013

  13. [21]

    Relocating underwater features au- tonomously using sonar-based slam.IEEE Journal of Oceanic Engineering, 38(3):500–513, 2013

    Maurice F Fallon, John Folkesson, Hunter McClelland, and John J Leonard. Relocating underwater features au- tonomously using sonar-based slam.IEEE Journal of Oceanic Engineering, 38(3):500–513, 2013. 1

    2013

  14. [22]

    Underwa- ter object detection: architectures and algorithms–a compre- hensive review.Multimedia Tools and Applications, 81(15): 20871–20916, 2022

    Sheezan Fayaz, Shabir A Parah, and GJ Qureshi. Underwa- ter object detection: architectures and algorithms–a compre- hensive review.Multimedia Tools and Applications, 81(15): 20871–20916, 2022. 1

    2022

  15. [23]

    Vision meets robotics: The kitti dataset.The in- ternational journal of robotics research, 32(11):1231–1237,

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.The in- ternational journal of robotics research, 32(11):1231–1237,

  16. [24]

    Digging into self-supervised monocular depth estimation

    Cl ´ement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 3828–3838,

  17. [25]

    Gonz ´alez-Sabbagh and Antonio Robles-Kelly

    Salma P. Gonz ´alez-Sabbagh and Antonio Robles-Kelly. A survey on underwater computer vision.ACM Comput. Surv., 55(13s), 2023. 2, 3

    2023

  18. [26]

    Uw-gan: Single-image depth estimation and image enhancement for underwater images.IEEE Transactions on Instrumentation and Measurement, 70:1–12, 2021

    Praful Hambarde, Subrahmanyam Murala, and Abhinav Dhall. Uw-gan: Single-image depth estimation and image enhancement for underwater images.IEEE Transactions on Instrumentation and Measurement, 70:1–12, 2021. 3, 6

    2021

  19. [27]

    Single image haze removal using dark channel prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 33(12):2341–2353,

    Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 33(12):2341–2353,

  20. [28]

    Distill any depth: Distillation creates a stronger monocular depth estimator, 2025

    Xiankang He, Dongyan Guo, Hongji Li, Ruibo Li, Ying Cui, and Chi Zhang. Distill any depth: Distillation creates a stronger monocular depth estimator, 2025. 2, 5

    2025

  21. [29]

    Teacher-student architecture for knowl- edge distillation: A survey, 2023

    Chengming Hu, Xuan Li, Dan Liu, Haolun Wu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowl- edge distillation: A survey, 2023. 2 9

    2023

  22. [30]

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geomet- ric foundation model for zero-shot metric depth and surface normal estimation.IEEE Transactions on Pattern Analysi...

    2024

  23. [31]

    Why warmup the learning rate? underlying mechanisms and improvements,

    Dayal Singh Kalra and Maissam Barkeshli. Why warmup the learning rate? underlying mechanisms and improvements,

  24. [32]

    Underwater optical-sonar image fusion systems.Sensors, 22(21):8445,

    Hong-Gi Kim, Jungmin Seo, and Soo Mee Kim. Underwater optical-sonar image fusion systems.Sensors, 22(21):8445,

  25. [33]

    Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution, 2022

    Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution, 2022. 3

    2022

  26. [34]

    An underwater image enhancement benchmark dataset and beyond.IEEE transac- tions on image processing, 29:4376–4389, 2019

    Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond.IEEE transac- tions on image processing, 29:4376–4389, 2019. 1

    2019

  27. [35]

    Sgdr: Stochastic gradient descent with warm restarts, 2017

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts, 2017. 4

    2017

  28. [36]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 4

    2019

  29. [37]

    A survey on vision-based uav navigation.Geo- spatial information science, 21(1):21–32, 2018

    Yuncheng Lu, Zhucun Xue, Gui-Song Xia, and Liangpei Zhang. A survey on vision-based uav navigation.Geo- spatial information science, 21(1):21–32, 2018. 1

    2018

  30. [38]

    John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J. Davison. Scenenet rgb-d: Can 5m synthetic im- ages beat generic imagenet pre-training on indoor segmenta- tion? InProceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2

    2017

  31. [39]

    Autonomous inspection using an underwater 3d lidar

    Dan McLeod, John Jacobson, Mark Hardy, and Carl Em- bry. Autonomous inspection using an underwater 3d lidar. In2013 OCEANS-San Diego, pages 1–8. IEEE, 2013. 1

    2013

  32. [40]

    Mobilevit: Light- weight, general-purpose, and mobile-friendly vision trans- former, 2022

    Sachin Mehta and Mohammad Rastegari. Mobilevit: Light- weight, general-purpose, and mobile-friendly vision trans- former, 2022. 2

    2022

  33. [41]

    Deep learning for monocular depth estimation: A review.Neuro- computing, 438:14–33, 2021

    Yue Ming, Xuyang Meng, Chunxiao Fan, and Hui Yu. Deep learning for monocular depth estimation: A review.Neuro- computing, 438:14–33, 2021. 1

    2021

  34. [42]

    Adjeroh, and Gi- anfranco Doretto

    Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gi- anfranco Doretto. Unified deep supervised domain adapta- tion and generalization. InProceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV), 2017. 3

    2017

  35. [43]

    Springer, 2021

    Sergey I Nikolenko et al.Synthetic data for deep learning. Springer, 2021. 1

    2021

  36. [44]

    A survey of structure from motion, 2017

    Onur Ozyesil, Vladislav V oroninski, Ronen Basri, and Amit Singer. A survey of structure from motion, 2017. 4

    2017

  37. [45]

    Autonomous mapping of underwater 3-d structures: From view planning to execution.IEEE Robotics and Au- tomation Letters, 3(3):1965–1971, 2018

    Narc ´ıs Palomeras, Natalia Hurt ´os, Marc Carreras, and Pere Ridao. Autonomous mapping of underwater 3-d structures: From view planning to execution.IEEE Robotics and Au- tomation Letters, 3(3):1965–1971, 2018. 1

    1965

  38. [46]

    Visual domain adaptation: A survey of recent advances.IEEE Signal Processing Magazine, 32(3):53–69,

    Vishal M Patel, Raghuraman Gopalan, Ruonan Li, and Rama Chellappa. Visual domain adaptation: A survey of recent advances.IEEE Signal Processing Magazine, 32(3):53–69,

  39. [47]

    Unidepthv2: Universal monocular metric depth estimation made simpler, 2025

    Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mat- tia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler, 2025. 2, 3, 5, 6, 7, 8

    2025

  40. [48]

    Z-splat: Z-axis gaussian splatting for camera-sonar fusion.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 2024

    Ziyuan Qu, Omkar Vengurlekar, Mohamad Qadri, Kevin Zhang, Michael Kaess, Christopher Metzler, Suren Jaya- suriya, and Adithya Pediredla. Z-splat: Z-axis gaussian splatting for camera-sonar fusion.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 2024. 1, 2

    2024

  41. [49]

    Svin2: An underwater slam system using sonar, visual, inertial, and depth sensor

    Sharmin Rahman, Alberto Quattrini Li, and Ioannis Rek- leitis. Svin2: An underwater slam system using sonar, visual, inertial, and depth sensor. In2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1861–1868. IEEE, 2019. 1

    2019

  42. [50]

    PhD thesis, ProQuest Dissertations Publishing,

    Yelena Randall and Tali Treibitz.FLSea: Under- water Visual-Inertial and Stereo-Vision Forward-Looking Datasets. PhD thesis, ProQuest Dissertations Publishing,

  43. [51]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InProceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 1, 2, 3, 4

    2021

  44. [52]

    Underwater image enhancement: a comprehensive re- view, recent trends, challenges and applications.Artificial Intelligence Review, 54:5413–5467, 2021

    Smitha Raveendran, Mukesh D Patil, and Gajanan K Biraj- dar. Underwater image enhancement: a comprehensive re- view, recent trends, challenges and applications.Artificial Intelligence Review, 54:5413–5467, 2021. 1, 6

    2021

  45. [53]

    Susskind

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021,

    2021

  46. [54]

    Mobilenetv2: Inverted residuals and linear bottlenecks, 2019

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 5, 6

    2019

  47. [55]

    Ashutosh Saxena, Jamie Schulte, and Andrew Y . Ng. Depth estimation using monocular and stereo cues. InProceedings of the 20th International Joint Conference on Artifical Intel- ligence, page 2197–2203, San Francisco, CA, USA, 2007. Morgan Kaufmann Publishers Inc. 1

    2007

  48. [56]

    Schechner and N

    Y .Y . Schechner and N. Karpel. Clear underwater vision. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages I–I, 2004. 4

    2004

  49. [57]

    Unsupervised low-light image enhancement by extracting structural similarity and color consistency.IEEE Signal Pro- cessing Letters, 29:997–1001, 2022

    Yangming Shi, Binquan Wang, Xiaopo Wu, and Ming Zhu. Unsupervised low-light image enhancement by extracting structural similarity and color consistency.IEEE Signal Pro- cessing Letters, 29:997–1001, 2022. 4

    2022

  50. [58]

    Solonenko and Curtis D

    Michael G. Solonenko and Curtis D. Mobley. Inherent op- tical properties of jerlov water types.Appl. Opt., 54(17): 5392–5401, 2015. 4, 8

    2015

  51. [59]

    How to train your vit? data, augmentation, and regularization in vision transformers.arXiv preprint arXiv:2106.10270, 2021

    Andreas Steiner, Alexander Kolesnikov, , Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers.arXiv preprint arXiv:2106.10270, 2021. 2 10

    2021 arXiv

  52. [60]

    Review of underwa- ter sensing technologies and applications.Sensors, 21(23): 7849, 2021

    Kai Sun, Weicheng Cui, and Chi Chen. Review of underwa- ter sensing technologies and applications.Sensors, 21(23): 7849, 2021. 1

    2021

  53. [61]

    Vi- sual slam algorithms: A survey from 2010 to 2016.IPSJ transactions on computer vision and applications, 9(1):16,

    Takafumi Taketomi, Hideaki Uchiyama, and Sei Ikeda. Vi- sual slam algorithms: A survey from 2010 to 2016.IPSJ transactions on computer vision and applications, 9(1):16,

    2010

  54. [62]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.CoRR, abs/1706.03762, 2017. 2

    2017 arXiv

  55. [63]

    Wagner, Nadieh Khalili, Raghav Sharma, Melanie Boxberg, Carsten Marr, Walter de Back, and Tingying Peng

    Sophia J. Wagner, Nadieh Khalili, Raghav Sharma, Melanie Boxberg, Carsten Marr, Walter de Back, and Tingying Peng. Structure-preserving multi-domain stain color augmentation using style-transfer with disentangled representations. In Medical Image Computing and Computer Assiste...

    2021

  56. [64]

    Underwa- ter localization and 3d mapping of submerged structures with a single-beam scanning sonar

    Jinkun Wang, Shi Bai, and Brendan Englot. Underwa- ter localization and 3d mapping of submerged structures with a single-beam scanning sonar. In2017 IEEE Inter- national Conference on Robotics and Automation (ICRA), pages 4898–4905. IEEE, 2017. 1

    2017

  57. [65]

    Domain adaptation for underwater im- age enhancement.IEEE Transactions on Image Processing, 32:1442–1457, 2023

    Zhengyong Wang, Liquan Shen, Mai Xu, Mei Yu, Kun Wang, and Yufei Lin. Domain adaptation for underwater im- age enhancement.IEEE Transactions on Image Processing, 32:1442–1457, 2023. 1, 3, 4, 5, 6

    2023

  58. [66]

    Junjie Wen, Jinqiang Cui, Zhenjun Zhao, Ruixin Yan, Zhi Gao, Lihua Dou, and Ben M. Chen. Syreanet: A physi- cally guided underwater image enhancement framework in- tegrating synthetic and real images. In2023 IEEE Inter- national Conference on Robotics and Automation (ICRA), pa...

    2023

  59. [67]

    Con- vnext v2: Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Con- vnext v2: Co-designing and scaling convnets with masked autoencoders. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1613...

    2023

  60. [68]

    O’Kane, and Ioannis Rekleitis

    Marios Xanthidis, Nare Karapetyan, Hunter Damron, Sharmin Rahman, James Johnson, Allison O’Connell, Ja- son M. O’Kane, and Ioannis Rekleitis. Navigation in the presence of obstacles for an agile autonomous underwater vehicle.arXiv preprint arXiv:1903.11750, 2020. 1

    1903 arXiv

  61. [69]

    A systematic review and analysis of deep learning-based underwater object detection.Neurocomput- ing, 527:204–232, 2023

    Shubo Xu, Minghua Zhang, Wei Song, Haibin Mei, Qi He, and Antonio Liotta. A systematic review and analysis of deep learning-based underwater object detection.Neurocomput- ing, 527:204–232, 2023. 1

    2023

  62. [70]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 1

    2024

  63. [71]

    Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024. 1, 2, 3, 4, 5, 6, 7, 8

    2024

  64. [72]

    Udepth: Fast monocular depth estimation for visually-guided underwater robots

    Boxiao Yu, Jiayi Wu, and Md Jahidul Islam. Udepth: Fast monocular depth estimation for visually-guided underwater robots. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3116–3123, 2023. 1, 2

    2023

  65. [73]

    Atlantis: En- abling underwater depth estimation with stable diffusion

    Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: En- abling underwater depth estimation with stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 11852–11861,

  66. [74]

    Survey on monocular metric depth estima- tion, 2025

    Jiuling Zhang. Survey on monocular metric depth estima- tion, 2025. 1

    2025

  67. [75]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3

    2023

  68. [76]

    An open-source, fiducial-based, un- derwater stereo visual-inertial localization method with re- fraction correction

    Pengfei Zhang, Zhengxing Wu, Jian Wang, Shihan Kong, Min Tan, and Junzhi Yu. An open-source, fiducial-based, un- derwater stereo visual-inertial localization method with re- fraction correction. InProceedings of the 2021 IEEE/RSJ International Conference on Intelligent Robots ...

    2021

  69. [77]

    Overview of underwater trans- mission characteristics of oceanic lidar.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:8144–8159, 2021

    Guoqing Zhou, Chenyang Li, Dianjun Zhang, Dequan Liu, Xiang Zhou, and Jie Zhan. Overview of underwater trans- mission characteristics of oceanic lidar.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:8144–8159, 2021. 1

    2021

  70. [78]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks.CoRR, abs/1703.10593,

    arXiv

  71. [79]

    Surrogate gap minimization improves sharpness-aware training.ICLR, 2022

    Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training.ICLR, 2022. 2

    2022

  72. [80]

    Underwater rgb- d camera based on binocular stereo vision.Acta Photonica Sin, 51:0404003, 2022

    SF Zhuang, Y Ji, DW Tu, and X Zhang. Underwater rgb- d camera based on binocular stereo vision.Acta Photonica Sin, 51:0404003, 2022. 1 11

    2022

  73. [2023]

    1, 2, 3, 4, 5, 6, 8

    PQDT - Global, ISBN: 9798379478926. 1, 2, 3, 4, 5, 6, 8

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.