Pith. sign in

REVIEW 4 major objections 4 minor 49 references

WAT3R: Feedforward Underwater 3D Reconstruction

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read WAT3R claims that underwater 3D reconstruction can be done feedforward—estimating dense point maps and camera poses from degraded underwater video in a single pass—by treating optical degradation as a geometry-constrained process rather tha

desk verdict A sensible, well-engineered feedforward underwater reconstruction paper with consistent but small gains; missing the most relevant baseline and overclaims pose. read the letter →

arxiv 2607.21023 v1 pith:JUP2W6X2 submitted 2026-07-23 cs.CV

Jiayi Xu , Jiahao Lu , Ziqiang Zheng , Yihao Tan , Yaolong Zhu , Yuan Liu , Sai-Kit Yeung This is my paper · ORCID
classification cs.CV
keywords underwater3Dreconstructionfeedforwardimageformationmodeldegradationadaptationcameraposeestimationmonoculardepthself-supervisedlearningpointmapprediction
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 tries to show that reliable underwater 3D reconstruction no longer needs slow per-scene optimization or clean images. WAT3R takes a terrestrial feedforward reconstruction model and adapts it to underwater inputs using the physics of light attenuation and backscattering. The adaptation happens through a lightweight neural module that predicts a restored 'clean' image, a head that estimates water parameters, and a two-stage training scheme: full supervision on synthetically degraded RGB-D data, then self-supervision on real underwater video. If the claim holds, underwater robots can estimate dense depth, point clouds, and camera trajectories from ordinary video, efficiently and scene-agnostically. The paper reports that this consistently outperforms existing feedforward baselines on multi-view and monocular depth benchmarks and on a real underwater pose benchmark.

What carries the argument

The load-bearing object is the Underwater Image Formation Model (UIFM), a per-channel equation I = J·e^(−β_att·z) + B∞·(1 − e^(−β_bs·z)) relating observed color I to true radiance J, depth z, and three global water parameters (attenuation, backscatter, background light). WAT3R uses it in three ways: to synthesize pseudo-underwater training pairs from terrestrial RGB-D data, as a soft reconstruction loss that ties the predicted clean image to the predicted depth, and as the target for a water-parameter prediction head. The architecture also includes a neural degradation-adaptation module that predicts a residual clean image, giving the network a degradation-invariant feature pathway.

What would settle it

A concrete test: take a real underwater sequence with independently measured or reconstructed ground-truth depth (e.g., from a structured-light or multi-camera rig), run WAT3R without any fine-tuning on that site, then retrain the same model on synthetic data with deliberately wrong water parameters (e.g., reversed Jerlov types). If depth accuracy does not drop when the water parameters are wrong, the UIFM adaptation is not doing the geometric work claimed.

Watch

Extended reading notes

Core claim

The central discovery is that the coupling between geometry and underwater radiance—the same physical model that causes color loss and scattering—can be repurposed as a training signal. WAT3R predicts pixel-aligned 3D point maps and relative camera poses in one forward pass, and jointly predicts a clean image via a residual degradation-adaptation module. The clean image is re-rendered through the underwater image formation model and compared with the observed input, giving a physics-based reconstruction loss; on real video, the clean images are used for self-supervised multi-view photometric consistency. The paper's key evidence is that this geometry-constrained adaptation, trained on synthe

Load-bearing premise

The claim rests on the assumption that terrestrial RGB-D images degraded by a simplified, globally homogeneous water formation model are close enough to real underwater optics that depth learned on them transfers to authentic scenes; the paper notes the model is simplified and that the self-supervised stage mainly refines poses, so real-scene depth accuracy is largely inherited from synthetic data.

Editorial extensions

If this is right

  • If correct, feedforward underwater reconstruction becomes a practical alternative to SfM-plus-optimization pipelines, enabling real-time dense reconstruction from underwater video without ground-truth geometry.
  • The same physics-guided adaptation recipe—synthetic degradation of labeled RGB-D data plus self-supervised photometric consistency on real data—could be transferred to other degraded domains such as haze, low light, or scattering media.
  • The clean images produced by the adaptation module are a byproduct that performs competitively with dedicated restoration methods on quality metrics, so a single network can serve reconstruction and restoration simultaneously.
  • Because the model is scene-agnostic and runs in one forward pass, it can become a drop-in component for underwater navigation, inspection, and mapping systems that previously relied on offline optimization.
  • The reported pose results suggest the self-supervised stage mainly improves rotational stability rather than absolute trajectory accuracy, indicating where the approach is strongest.

Reading between the lines

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

  • Beyond the paper: a reader should expect the depth advantage to be partly an artifact of training on synthetic data generated from the same formation model used in the loss; the real test is whether the trend holds with measured inherent optical properties and independent field data.
  • Beyond the paper: if the simplified homogeneous-IOP assumption is the bottleneck, extending the formation model to spatially varying attenuation and backscatter fields would be a natural next step; the paper's own appendix notes plankton patchiness and stratification as unmodeled effects.
  • Beyond the paper: the restoration-versus-geometry coupling suggests a testable extension—deliberately corrupting the water-parameter head should degrade depth accuracy only if the geometry is truly using the physics signal, not just image statistics.
  • Beyond the paper: the pose evaluation on a small self-collected set with structure-from-motion pseudo-ground truth would benefit from independent trajectory validation against inertial or acoustic ground truth before taking the pose numbers at face value.
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 / 4 minor

Summary. WAT3R is a feedforward framework for underwater 3D reconstruction built on a Pi3-style transformer backbone. It uses a two-stage training pipeline: (i) supervised adaptation on pseudo-underwater versions of terrestrial RGB-D datasets, generated with the simplified UIFM of Eq. (1) with Jerlov-class IOPs, and (ii) self-supervised refinement on real underwater videos via masked photometric consistency (Eqs. 6–7), which the paper states is used mainly to optimize camera pose. The paper reports multi-view depth on FLSea/SQUID, monocular depth on USOD10K, pose on a self-collected dataset, and auxiliary restoration results in the appendix.

Significance. If the reported gains are robust, WAT3R would be a practical feedforward underwater reconstruction method and a useful demonstration that a soft UIFM constraint can improve geometry. The method is clearly described, and the supplementary provides unusually detailed IOP-sampling and training-loss information, which is a strength for reproducibility. However, the evidence currently does not support the abstract's blanket claim of consistent state-of-the-art performance: several comparisons are within noise, the most relevant baseline is absent, and depth generalization rests on the synthetic premise.

major comments (4)
  1. [Abstract; Tables 2 and 3] The abstract claims consistent superiority on camera pose estimation, but Table 3 shows WAT3R is not best on ATE (0.2321 vs DA3's 0.2025) or RPEtrans (0.0721 vs Pi3's 0.0680); only RPErot is best. Similarly, Table 2 shows DA3 has a higher δ<1.25 (0.5330 vs 0.4916), and Table 1 (scale-and-shift, SQUID) shows MapAnything has a lower AbsRel (0.1332 vs 0.1369). The claim should be narrowed to 'competitive or best on most metrics' or the abstract revised to match the tables.
  2. [Section 2; Tables 1–3] SeaVGGT is identified in Related Work as a feedforward self-supervised underwater geometry method, yet it does not appear in any experimental comparison. Since SeaVGGT is the most relevant feedforward underwater baseline, its omission prevents the 'state-of-the-art' claim from being substantiated. A comparison should be added.
  3. [Tables 1–2] Several headline margins are extremely small: FLSea scale-only AbsRel is 0.0725 for WAT3R vs 0.0728 for TRUDepth, and scale-and-shift is 0.0693 vs 0.0698. No standard deviations, repeated-run statistics, or significance tests are reported. Given the 'consistently outperforms' claim, the lack of any uncertainty quantification is a load-bearing gap.
  4. [Section 3.3; Eq. (1)] Depth on real underwater images is almost entirely inherited from the synthetic adaptation stage, because the real-data self-supervised stage explicitly focuses on pose only ('we focus on optimizing the camera pose using the clean images'). If Eq. (1)'s simplified formation model is unrepresentative of real scattering/absorption coupling, the reported depth gains could be an artifact of the synthetic generator. The paper should provide an analysis separating synthetic-only depth performance from final depth performance, or otherwise validate that the synthetic depth prior transfers to real geometry.
minor comments (4)
  1. [Section 3.3, Eq. (5)] Lrecon alone admits a trivial solution (J = I, β = 0). In the supervised stage this is prevented by ground-truth depth/water-parameter losses, but the paper should state this explicitly to avoid the appearance of an unsupervised physical-consistency constraint that could be degenerate.
  2. [Section 4.4, Figure 6 caption] There are typos: 'qualitive' and 'quantative' should be 'qualitative' and 'quantitative'. The figure caption also has a stray closing parenthesis.
  3. [Section 4.3 / Appendix E.1] The pose benchmark is self-collected from public videos with COLMAP poses as ground truth. Please report the number of frames per sequence, total trajectory length, and any quality filtering applied to the COLMAP poses, since underwater COLMAP can be unreliable.
  4. [Appendix C] The restoration results are auxiliary to the 3D-reconstruction claim. The UIQM analysis is interesting, but the section feels long relative to the main paper; consider compressing and clearly labeling it as a side observation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UIFM is used as a synthetic-data generator and soft regularizer, not as the source of the evaluated geometry claims.

full rationale

The paper's central claim is empirical: WAT3R is trained on UIFM-degraded synthetic RGB-D data and then evaluated on real underwater datasets. The UIFM appears twice — as a data generator and as the soft reconstruction loss L_recon — but neither occurrence makes the target result an identity. The paper explicitly disclaims a hard constraint: "Equation (1) is a simplified and imperfect underwater imaging model. Thus, our method does not put Equation (1) as a hard constraint in the geometry estimation. Instead, we apply it to construct a pseudo dataset for the domain adaptation. Then, we formulate a loss function based on the equation, which acts like a soft regularization." In the supervised stage, depth and pose are supervised directly by ground truth from terrestrial RGB-D datasets, and the self-supervised stage explicitly "focus[es] on optimizing the camera pose using the clean images," so real-world depth is not derived by inverting Eq. (1) on the test set. The degenerate solution of L_recon (J≈I, water parameters≈0) is counteracted by the supervised losses L_points, L_cam, L_water, and L_J. Evaluation on FLSea, SQUID, and USOD10K uses external or pseudo ground truth, not outputs of the paper's generator. The self-citations present, e.g., Align3r [20] and UVEB [27], are related-work/data-source references, not load-bearing uniqueness theorems or forced ansätze. The acknowledged limitations — simplified UIFM, locally homogeneous IOPs, USOD10K pseudo-depth, and turbid-water failures — are external-validity concerns, not evidence that any prediction reduces to the training inputs by construction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The paper makes no first-principles derivation; its 'physics' enters as a generator and a soft loss. Hence the ledger is dominated by domain assumptions about synthetic-to-real transfer and evaluation, not by fitted constants. The known water parameters are estimated per scene by a learned head, so they are not unaccounted free constants.

free parameters (3)
  • Loss weights = λ_J=0.5, λ_recon=0.3, λ_points=1.0, λ_cam=0.2, λ_water=0.3; L1/perceptual/SSIM weights 0.8/0.1/0.2; photometric SSIM/L1
    Chosen by hand (Supp. B.2) and not ablated for sensitivity; the reported ablation only turns L_recon off.
  • Global scale factor s-hat = per-sample during supervised point loss (Eq. 17)
    Aligns predicted point maps to GT to resolve monocular scale ambiguity; standard practice but a fitted quantity in the supervision.
  • Pose-only self-supervised objective = not applicable (modeling choice)
    Deciding to optimize only pose during real-data fine-tuning is a hand-made choice that determines how much real-domain depth adaptation occurs; no depth fine-tuning on real data is reported.
assumptions (6)
  • domain assumption Simplified UIFM (Eq. 1) with globally homogeneous IOPs is an adequate model for generating training data and supervising geometry.
    Section 3.1 ('simplified and imperfect'); used for synthetic data and L_recon. If real underwater optics deviate, depth supervision is biased.
  • domain assumption Jerlov-derived IOP samples from WOOD, filtered to top 10 m and six water types, represent the optical environments of FLSea, SQUID, USOD10K, and the self-collected videos.
    Supp. D describes campaign grouping and homogeneity assumption; the paper notes geographic limitations.
  • domain assumption A terrestrial feedforward 3D model (Pi3) retains and transfers its geometry prior after underwater adaptation.
    Section 3.2 and Supp. B: Pi3 backbone initialization; if fine-tuning destroys the prior, the approach collapses.
  • domain assumption Clean-image photometric consistency across views, with minimum-reprojection auto-masking, provides a valid self-supervised signal for real underwater sequences.
    Section 3.3 Eq. (7) and Supp. B.3; inherited from Godard et al. [15]; assumes Lambertian/static scenes except masked pixels.
  • domain assumption COLMAP reconstructions of self-collected underwater videos provide accurate ground-truth camera poses.
    Supp. E.1; degraded underwater images can break SfM matching, so the ATE/RPE evaluation inherits COLMAP failures.
  • domain assumption DPT pseudo-depth in USOD10K is a meaningful reference for comparing monocular depth models.
    Section 4.2 acknowledges pseudo-depth inaccuracies; using it as the benchmark can favor models whose inductive biases resemble DPT.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WAT3R: Feedforward Underwater 3D Reconstruction." pith.science (2026). https://pith.science/paper/JUP2W6X2

@misc{pith2026260721023,
  author       = {Pith},
  title        = {Pith review of: WAT3R: Feedforward Underwater 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUP2W6X2}},
  note         = {Machine review of arXiv:2607.21023}
}
read the original abstract

Reliable feedforward underwater 3D reconstruction remains challenging due to severe light attenuation and backscattering, which degrade visual quality and disrupt feature consistency across views, leading to inaccurate multi-view geometry. To address this issue, we propose WAT3R, a feed-forward framework for reconstructing 3D scenes directly from underwater images. By leveraging degradation adaptation as a geometry-constrained process, WAT3R integrates a lightweight neural adaptation module to flexibly account for these underwater imaging effects, thereby improving multi-view reconstruction quality. Implemented in a single forward pass, WAT3R directly and efficiently outputs pixel-aligned 3D point maps and camera poses from underwater videos, allowing a high-quality underwater 3D reconstruction. Experiments conducted on the FLSea, SQUID, and USOD10K datasets show that our method consistently outperforms state-of-the-art approaches on 3D reconstruction tasks, including multi-view/monocular depth estimation and camera pose estimation.

Figures

Figures reproduced from arXiv: 2607.21023 by the authors.

Figure 1
Figure 1. Given degraded underwater images (left), [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the WAT3R framework. Our method integrates a degradation adaptation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed learning pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Simulation of various underwater optical environments. Based on the clean in￾air image and metric depth map (left), we synthesize physically plausible observations across different Jerlov water types. Synthetic Data Generation. To overcome the scarcity of real-world un…
Figure 5
Figure 5. Figure 5: Qualitative comparison of depth estimation results on underwater scenes. Colors indicate [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of 3D point cloud reconstruction on underwater scenes. Compared [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 2 linked inside Pith

  1. [1]

    Rekleitis

    Sharmin Rahman, Alberto Quattrini Li, and Ioannis M. Rekleitis. Svin2: Sonar visual-inertial SLAM with loop closure for underwater navigation.CoRR, 2018

    2018

  2. [2]

    Pengyu Wang, Hin Wang Lin, Jialu Li, Jiankun Wang, Ling Shi, and Max Q.-H. Meng. Pierguard: A planning framework for underwater robotic inspection of coastal piers.IEEE Trans Autom. Sci. Eng., 22:15941–15952, 2025

    2025

  3. [3]

    Derya Akkaynak, Tali Treibitz, Tom Shlesinger, Yossi Loya, Raz Tamir, and David Iluz. What is the space of attenuation coefficients in underwater computer vision? In2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 568–577. IEEE Computer Society, 2017

    2017

  4. [4]

    Seathru-nerf: Neural radiance fields in scattering media

    Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru-nerf: Neural radiance fields in scattering media. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 56–65, 2023

    2023

  5. [5]

    Waternerf: Neural radiance fields for underwater scenes

    Advaith Venkatramanan Sethuraman, Manikandasriram Srinivasan Ramanagopal, and Kather- ine A Skinner. Waternerf: Neural radiance fields for underwater scenes. InOCEANS 2023- MTS/IEEE US Gulf Coast, pages 1–7. IEEE, 2023

    2023

  6. [6]

    Waterhe-nerf: Water-ray tracing neural radiance fields for underwater scene reconstruction, 2024

    Jingchun Zhou, Tianyu Liang, Dehuan Zhang, and Zongxin He. Waterhe-nerf: Water-ray tracing neural radiance fields for underwater scene reconstruction, 2024

    2024

  7. [7]

    Watersplatting: Fast underwater 3d scene reconstruction using gaussian splatting

    Huapeng Li, Wenxuan Song, Tianao Xu, Alexandre Elsig, and Jonas Kulhanek. Watersplatting: Fast underwater 3d scene reconstruction using gaussian splatting. In2025 International Conference on 3D Vision (3DV), pages 969–978. IEEE, 2025

    2025

  8. [8]

    Seasplat: Representing underwater scenes with 3d gaussian splatting and a physically grounded image formation model

    Daniel Yang, John J Leonard, and Yogesh Girdhar. Seasplat: Representing underwater scenes with 3d gaussian splatting and a physically grounded image formation model. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 7632–7638. IEEE, 2025

    2025

Show all 49 references
  1. [9]

    Waterclear-gs: Optical-aware gaussian splatting for underwater reconstruction and restoration, 2026

    Xinrui Zhang, Yufeng Wang, Shuangkang Fang, Zesheng Wang, Dacheng Qi, and Wenrui Ding. Waterclear-gs: Optical-aware gaussian splatting for underwater reconstruction and restoration, 2026

    2026

  2. [10]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025

    2025

  3. [11]

    pi3: Permutation-equivariant visual geometry learning.arXiv preprint arXiv:2507.13347, 2025

    Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. pi3: Permutation-equivariant visual geometry learning.arXiv preprint arXiv:2507.13347, 2025

    2025 arXiv

  4. [12]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y . Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views, 2025

    2025

  5. [13]

    Mapanything: Universal feed-forward metric 3d reconstruction, 2026

    Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, Jonathon Luiten, Manuel Lopez-Antequera, Samuel Rota Bulò, Christian Richardt, Deva Ramanan, Sebastian Scherer, and Peter...

    2026

  6. [14]

    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. 10

    2018

  7. [15]

    Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J. Brostow. Digging into self-supervised monocular depth prediction. October 2019

    2019

  8. [16]

    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. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1682–1691, 2019

    2019

  9. [17]

    Neural underwater scene representation

    Yunkai Tang, Chengxuan Zhu, Renjie Wan, Chao Xu, and Boxin Shi. Neural underwater scene representation. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11780–11789, 2024

    2024

  10. [18]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jérôme Revaud. Dust3r: Geometric 3d vision made easy. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20697–20709. IEEE, 2024

    2024

  11. [19]

    Monst3r: A simple approach for estimating geometry in the presence of motion.arXiv preprint arXiv:2410.03825, 2024

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion.arXiv preprint arXiv:2410.03825, 2024

    2024 arXiv

  12. [20]

    Align3r: Aligned monocular depth estimation for dynamic videos

    Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22820–22830, 2025

    2025

  13. [21]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass

    Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21...

    2025

  14. [22]

    Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views

    Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. InProceedings of the Computer Vision and Pattern Recognition...

    2025

  15. [23]

    Underwater visual geometry estimation with self-supervised prototype-graph modulation, 2026

    Linqing Zhao, Wenzhao Zheng, Zelan Zhu, Haibin Yan, and Jiwen Lu. Underwater visual geometry estimation with self-supervised prototype-graph modulation, 2026

    2026

  16. [24]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. CoRR, 2021

    2021

  17. [25]

    Williamson and Richard C

    Craig A. Williamson and Richard C. Hollins. Measured iops of jerlov water types.Appl. Opt., 61(33):9951–9961, Nov 2022

    2022

  18. [26]

    A rapid scene depth estima- tion model based on underwater light attenuation prior for underwater image restoration

    Wei Song, Yan Wang, Dongmei Huang, and Dian Tjondronegoro. A rapid scene depth estima- tion model based on underwater light attenuation prior for underwater image restoration. In Advances in Multimedia Information Processing – PCM 2018: 19th Pacific-Rim Conference on Multimedi...

    2018

  19. [27]

    Uveb: A large-scale benchmark and baseline towards real-world underwater video enhancement

    Yaofeng Xie, Lingwei Kong, Kai Chen, Ziqiang Zheng, Xiao Yu, Zhibin Yu, and Bing Zheng. Uveb: A large-scale benchmark and baseline towards real-world underwater video enhancement. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22358–22367, 2024

    2024

  20. [28]

    Flsea: Underwater visual-inertial and stereo-vision forward- looking datasets, 2023

    Yelena Randall and Tali Treibitz. Flsea: Underwater visual-inertial and stereo-vision forward- looking datasets, 2023

    2023

  21. [29]

    Underwater single image color restoration using haze-lines and a new quantitative dataset.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020

    Dana Berman, Deborah Levy, Shai Avidan, and Tali Treibitz. Underwater single image color restoration using haze-lines and a new quantitative dataset.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 11

    2020

  22. [30]

    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. InIEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023

    2023

  23. [31]

    Metrically scaled monocular depth estima- tion through sparse priors for underwater robots, 2023

    Luca Ebner, Gideon Billings, and Stefan Williams. Metrically scaled monocular depth estima- tion through sparse priors for underwater robots, 2023

    2023

  24. [32]

    Usod10k: A new benchmark dataset for underwater salient object detection.IEEE Transactions on Image Processing, 34:1602–1615, 2025

    Lin Hong, Xin Wang, Gan Zhang, and Ming Zhao. Usod10k: A new benchmark dataset for underwater salient object detection.IEEE Transactions on Image Processing, 34:1602–1615, 2025

    2025

  25. [33]

    Virtual kitti 2, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020

    2020

  26. [34]

    Tartanair: A dataset to push the limits of visual slam

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. 2020

    2020

  27. [35]

    Deep- mvs: Learning multi-view stereopsis

    Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deep- mvs: Learning multi-view stereopsis. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  28. [36]

    Nisha Varghese, Ashish Kumar, and A. N. Rajagopalan. Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12248–12258, October 2023

    2023

  29. [37]

    Osmosis: RGBD diffusion prior for underwater image restoration

    Opher Bar Nathan, Deborah Levy, Tali Treibitz, and Dan Rosenbaum. Osmosis: RGBD diffusion prior for underwater image restoration. InComputer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXII, volume 15120 ofLectur...

    2024

  30. [38]

    Indoor segmentation and support inference from rgbd images

    Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. InECCV, 2012

    2012

  31. [39]

    An underwater image enhancement benchmark dataset and beyond, 2019

    Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond, 2019

    2019

  32. [40]

    Human-visual-system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2016

    Karen Panetta, Chen Gao, and Sos Agaian. Human-visual-system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2016

    2016

  33. [41]

    Musiq: Multi-scale image quality transformer, 2021

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer, 2021

    2021

  34. [42]

    Jiayi Wu, Tianfu Wang, Md Abu Bakr Siddique, Md Jahidul Islam, Cornelia Fermuller, Yiannis Aloimonos, and Christopher A. Metzler. Single-step latent diffusion for underwater image restoration, 2025

    2025

  35. [43]

    Histoformer: Histogram-based transformer for efficient underwater image enhancement.IEEE Journal of Oceanic Engineering, 50(1):164–177, 2025

    Yan-Tsung Peng, Yen-Rong Chen, Guan-Rong Chen, and Chun-Jung Liao. Histoformer: Histogram-based transformer for efficient underwater image enhancement.IEEE Journal of Oceanic Engineering, 50(1):164–177, 2025

    2025

  36. [44]

    Unsupervised underwater image restoration: From a homology perspective

    Zhenqi Fu, Huangxing Lin, Yan Yang, Shu Chai, Liyan Sun, Yue Huang, and Xinghao Ding. Unsupervised underwater image restoration: From a homology perspective. InThirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applicatio...

    2022

  37. [45]

    U-shape transformer for underwater image enhance- ment.IEEE Trans

    Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhance- ment.IEEE Trans. Image Process., 32:3066–3079, 2023

    2023

  38. [46]

    Liteenhancenet: A lightweight network for real-time single underwater image enhancement.Expert Syst

    Song Zhang, Shili Zhao, Dong An, Daoliang Li, and Ran Zhao. Liteenhancenet: A lightweight network for real-time single underwater image enhancement.Expert Syst. Appl., 240:122546, 2024. 12

    2024

  39. [47]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes Lutz Schönberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. InEuropean Conference on Computer Vision (ECCV), 2016

    2016

  40. [48]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  41. [49]

    campaigns

    Carlos Campos, Richard Elvira, Juan J. Gómez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6):1874–1890, 2021. 13 A Overview This supplementary material prov...

    2021

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.