Uncertainty Aware Mapping for Vision-Based Underwater Robots

Abhimanyu Bhowmik; Kostas Alexis; Madhushree Sannigrahi; Martin Ludvigsen; Mohit Singh

REVIEW 3 major objections 5 minor 12 references

Uncertainty Aware Mapping for Vision-Based Underwater Robots

T0 review · 3 major / 5 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 Stereo confidence now colors the 3D maps of underwater robots.

desk verdict A useful engineering integration of stereo confidence into Voxblox, but the printed confidence equation is self-similarity rather than left-right matching, and the change to the weight update alters what the confidence colors actually mean. read the letter →

arxiv 2507.10991 v1 pith:R5DIQZQL submitted 2025-07-15 cs.RO

Abhimanyu Bhowmik , Mohit Singh , Madhushree Sannigrahi , Martin Ludvigsen , Kostas Alexis This is my paper · ORCID

classification cs.RO

keywords underwaterroboticsvision-basedmappingdepthconfidencestereoestimationTSDFVoxbloxuncertaintyRAFT-Stereo

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 proposes a way to make 3D maps built by vision-based underwater robots carry a measure of how much each depth estimate is to be trusted. It uses RAFT-Stereo, a learning-based stereo depth model, to produce a per-pixel confidence score alongside the depth, and injects that confidence into a voxel-based mapping framework by setting each voxel's weight from confidence instead of from distance. The paper also replaces the framework's weight accumulation rule with an averaging rule, so that re-observing a surface does not inflate confidence unless the view itself changes. Experiments with a custom ROV in a pool and in a Trondheim fjord pier show that texture-rich surfaces receive higher confidence and that a second lap leaves the confidence map essentially unchanged. The result is a map whose colors tell operators and planners which reconstructed surfaces are reliable and which need another look.

What carries the argument

The carrying mechanism is a confidence-weighted TSDF update. A truncated signed distance field stores, in each voxel, a signed distance to the nearest surface and a weight; Voxblox normally weights a measurement by an inverse-square-of-distance model. The paper substitutes a weight that is a piecewise-constant function of the stereo confidence at the relevant pixel (and at the edge pixel), then replaces Voxblox's additive weight accumulation with the averaging rule $\Omega_{k+1} = \min((\Omega_k + \omega)/2, \Omega_{\max})$, so that confidence never grows merely by re-scanning a surface from the same viewpoint. The confidence itself comes from the cosine similarity between the left feature map and its disparity-shifted copy, which is used both to filter the depth cloud and to color the final map.

What would settle it

Render the same left image paired with two very different right images; if the paper's confidence map is identical in both cases, then the confidence never used the right image and cannot represent stereo matching confidence.

Watch

Extended reading notes

Core claim

The central claim is that stereo depth confidence can be made a first-class quantity inside a truncated signed distance field (TSDF) map. The authors compute a per-pixel confidence as the cosine similarity between the left image's feature vector and its disparity-shifted version, threshold the depth map by that confidence, and discard low-confidence points. In the Voxblox mapping framework, they initialize each voxel's weight from the confidence value rather than from the sensor-distance model, and they change the weight update from accumulation to a running average. The outcome is a voxel map where color encodes how much the stereo matcher trusted the depth that placed each surface, and re-visiting an area does not raise confidence unless the viewing geometry changes substantially.

Load-bearing premise

The whole confidence pipeline depends on the claim that shifting the left image's feature map sideways and comparing it to itself measures how well the left and right images match; if that self-comparison is not actually a stereo comparison, the confidence values do not mean what the paper says.

Editorial extensions

If this is right

The confidence-colored voxel map gives an immediate visual indication of which parts of an underwater reconstruction are trustworthy after a single pass.
Repeating a trajectory without changing viewpoint will no longer artificially raise map confidence, so the map reflects what the view actually resolved.
Texture-rich surfaces can be marked as more confident than texture-poor surfaces even when the texture-poor surface is closer, matching the failure modes of learning-based stereo.
Low-confidence depth points are filtered before fusion, so blank regions in the map pinpoint where the stereo matcher failed rather than where geometry is genuinely missing.
Because confidence is stored per voxel, the same map can serve as a cost layer for planning re-inspection routes that seek out low-confidence areas.

Reading between the lines

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

A true left–right consistency measure, comparing the left feature map with the shifted right feature map, would likely remove the systematic blank strip the paper observes on the left side of its confidence maps, and would make the confidence scores more interpretable.
The same confidence-weighting formalism could be transferred to other learned depth estimators by swapping in their feature maps, making the approach independent of RAFT-Stereo.
An online planner could treat the averaged weight as a measure of observation novelty: a viewpoint that does not change the weight is informationally redundant, which gives a cheap criterion for next-best-view selection.
Weighting each new measurement by its per-pixel confidence during TSDF fusion would let high-confidence views dominate the reconstruction, which is a natural extension of the averaging rule and untested in the paper.

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.

Desk Editor's Note

read the letter

Short version: this is a plausible engineering integration with real underwater field data, but the central confidence equation as printed does not measure what the paper claims, and the evaluation is entirely qualitative. If Eq. (1) is a typo, the text needs to be fixed; as written, the method is not reproducible.

The genuinely new piece is combining RAFT-Stereo's feature-vector cosine similarity as a confidence measure, using that confidence to initialize Voxblox weights, and changing the weight update from accumulation to a running average. That combination is not in the cited prior work. The field experiments in a test tank and a Trondheim fjord are real, and the figures show the expected behavior: texture-rich regions get higher confidence, dark blank areas get filtered out, and repeated laps do not automatically raise the displayed confidence. The authors are honest that the results are qualitative, which is appropriate for a workshop paper.

The soft spots are real and load-bearing. First, Eq. (1) in Section II.A computes the cosine similarity between the left feature vector and its own disparity-shifted version; there is no appearance of the right image feature vector anywhere. That is self-similarity along the epipolar line, not a stereo matching cost. A textureless wall or repetitive pattern can score high even when no reliable correspondence exists. Every downstream map color and filtering step inherits this. If it's a typo, the methodology section simply does not match the implementation, and the paper cannot be reproduced as written.

Second, Eq. (12) replaces Voxblox's cumulative weight update with an average. After N observations of the same surface, the total weight tends to 1 rather than growing with N. That may be intentional—they want revisiting not to increase confidence—but it means the displayed voxel confidence no longer represents accumulated evidence. The paper should state that trade-off explicitly.

Third, there is no quantitative validation: no error metrics against ground truth, no comparison to unmodified Voxblox, no code or data release. Given the confidence filter directly affects the point cloud, it would be straightforward to report something like fraction of valid depths or surface error.

Who is this for? People building vision-based inspection and mapping systems for constrained underwater environments. The idea is worth engaging with, and the field data is useful. But the confidence equation and weight-update semantics need to be corrected and clarified before the maps can be trusted to encode stereo matching uncertainty. I would not desk reject this; it deserves a serious referee. With a corrected equation, a quantitative section, and released artifacts, it could be a solid contribution.

Referee Report

3 major / 5 minor

Summary. The paper proposes a vision-based uncertainty-aware mapping pipeline for underwater robots. Depth and per-pixel confidence are obtained from RAFT-Stereo, and the confidence is injected into a Voxblox TSDF map by initializing voxel weights from confidence values and replacing Voxblox's accumulating weight update with an averaging update. The system is demonstrated qualitatively in a pool and in a Trondheim fjord pier, with the central claims being that texture-rich regions show higher reconstructed confidence and that revisiting an area does not automatically increase confidence unless the viewing frustum changes.

Significance. If the central methodology were sound, the paper would address a practical need: enabling low-cost vision-based inspection of confined underwater spaces with an explicit representation of stereo-matching uncertainty in the map. The combination of a modern learned stereo matcher (RAFT-Stereo) with a standard TSDF mapping framework (Voxblox) is attractive, and the proposed weight-averaging update, if properly justified, would be a simple way to prevent unbounded weight growth. However, the paper currently lacks machine-checked or quantitative evidence, and two load-bearing technical points—the confidence equation and the modified weight update—undermine the interpretation of the results as stereo-matching confidence and accumulated evidence.

major comments (3)

[II.C, Eq. (12)] As written, the cosine-similarity confidence is computed between the left feature vector g_theta(I_L) and its disparity-shifted version D_{L->R}^S(g_theta(I_L)), with no appearance of the right feature vector g_theta(I_R). This is not a stereo matching cost: shifting a feature map and comparing it with itself measures self-similarity along the epipolar line, not agreement between the two cameras. Textureless regions or repetitive patterns can therefore score high even when there is no reliable correspondence, and occluded pixels are not penalized. The paper's stated goal of estimating stereo depth confidence requires a comparison of corresponding left and right features. If this is a typo and the intended expression involves g_theta(I_R), then the printed methodology does not match the implementation and the paper is not reproducible as written; if it is not a typo, the confidence maps shown in Figs. 3–5 do not represent stereo matching confidence. Because every downstream filtering and coloring step inherits this quantity, the central claim of the paper cannot be accepted without correcting the equation and re-deriving the experimental conclusions.
[II.C, Eq. (12)] Replacing Voxblox's weight accumulation rule Omega_{k+1} = min(Omega_k + omega, Omega_max) with the averaging rule Omega_{k+1} = min((Omega_k + omega)/2, Omega_max) changes the semantics of the TSDF weight. With repeated observations of the same surface and omega near 1, the total weight tends to 1 rather than growing with the number of measurements, so the displayed voxel weight no longer represents accumulated evidence. The paper's conclusion that 'revisiting a place doesn't guarantee confidence increment' is therefore not an empirical finding but a direct consequence of this averaging update. The authors should either provide a formal argument that the averaging rule preserves the meaning of the TSDF weight in Eqs. (9)–(11), or present quantitative evidence that reconstruction accuracy is maintained with this modification.
[III.A] All results are qualitative: no quantitative evaluation of the confidence measure against actual depth error, no comparison with other confidence methods, no ablation of the threshold Cmin or of the proposed weight initialization versus Voxblox's default weighting, and no metric such as reconstruction error or map consistency. Moreover, the observation that texture-rich surfaces show higher confidence is partly circular, since cosine similarity between a feature map and a shifted version of itself will naturally be higher when features are distinct. A concrete test, such as binning estimated depth errors against confidence values or comparing the TSDF weights against actual measurement noise, is needed to support the claim that the proposed confidence reflects true matching uncertainty.

minor comments (5)

[II.C, Eq. (7)] The piecewise definition in Eq. (6) is typeset ambiguously; the conditions and the fraction (rho+tau)/(tau-eta) should be parenthesized explicitly so that the intended weighting is unambiguous.
[II.A] The symbols C1 and C2 are defined only in prose; please specify how C2 (the confidence of the 'edge' pixel) is computed and how it is distinguished from C1 at the voxel location.
[III.A, Fig. 4] The text states there is a 'skew towards the left' and then refers to a 'shift to the right'; this is confusing and should be clarified with a precise description of the direction of the disparity shift and its effect on the blank strip.
[References] Reference [10] appears to be the same work as [5], and the dataset link mentioned in Section III is not given; please disambiguate and add a direct URL for the dataset.
[General] The manuscript would benefit from a statement of code and data availability, especially because the RAFT-Stereo and Voxblox components are standard but the exact modifications and parameter values are needed for reproducibility.

Circularity Check

2 steps flagged · score 4.0 of 10

Two qualitative findings reduce to the paper's own definitions: texture-dependent confidence is built into the feature-similarity cost, and the no-confidence-growth-on-revisit observation is encoded in the averaging weight update.

self definitional [Section II.A, unnumbered cosine-similarity equation; Sections III.A-B and IV texture-dependence observations]
"The cosine similarity between the corresponding pixel's feature vectors g_theta(I_L) and the disparity-shifted feature vectors D^S_{L->R}(g_theta(I_L)) is computed as: CS_{L->R} = ... This cosine similarity produces a confidence map which is used for uncertainty estimation and depth point cloud filtering."

Confidence is defined, not independently measured, as the cosine similarity of the left feature map to its own disparity-shifted copy; the right-image feature vector g_theta(I_R) never appears in the formula. The later 'finding' that texture-rich regions have higher confidence (Sec. III.A: 'the fjord bed is texture-rich, which results in higher depth confidence') and that 'depth confidence ... [depends] on the textures' (Sec. III.B) is therefore a property of the chosen statistic, following from the definition rather than from an external uncertainty benchmark. The claimed 'dependence of depth confidence over the surface textures' is an input assumption of the confidence definition, so the validation is self-definitional.
self definitional [Section II.C, Eq. (12); Section IV conclusion about revisits; Fig. 5]
"To counter this, the equation has been modified to 12, which updates the weight with the average of current and previous weights. ... We also demonstrated that revisiting a place doesn't guarantee confidence increment unless the view frustum is changing significantly at each lap."

Equation (12) reads Ω_{k+1}=min((Ω_k+ω)/2, Ω_max). For repeated observations with the same confidence ω, this recurrence contracts toward min(ω, Ω_max), so the accumulated weight does not grow with the number of revisits. The authors introduced this rule precisely 'to counter' the old accumulation (Eq. 5). Consequently, the 'demonstrated' conclusion that revisiting does not guarantee a confidence increase unless the view changes is an analytic consequence of the update rule they designed, not an independent empirical result. The Fig. 5 two-lap comparison validates the design by the design itself.

full rationale

The two flagged steps are partial self-definitional reductions: the confidence map's texture dependence is inherent in the feature-similarity definition, and the no-accumulation-on-revisit behavior is built into the averaging update. However, the paper's central contribution—using that confidence to initialize Voxblox weights and integrating it into a TSDF mapping pipeline—does not reduce to these definitions; it is implemented and exercised on real underwater data. There is no load-bearing self-citation chain: ReAqROVIO [5,10] supplies odometry but is not used to justify the confidence or update claims. A separate correctness risk, not scored as circularity, is that the confidence equation compares the left feature map with its own shifted copy rather than with the right feature map, so the quantity may not be a stereo matching confidence as written. Overall this is partial circularity, not full.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on four assumptions: the validity of a new cosine-similarity confidence signal, the neglect of pose uncertainty, an ad hoc averaging weight update, and the transfer of a terrestrial RGB-D sensor model to underwater stereo. None are validated quantitatively in the paper.

free parameters (2)

Cmin (minimum confidence threshold) = not specified
Pixels with confidence below Cmin are discarded, producing blank patches in the depth map (Fig. 4d). The threshold is chosen by hand and not reported.
Voxel size mu and truncation distance tau = 4 mu = not specified
TSDF map resolution parameters; tau is set relative to mu as in Voxblox. They are user-selected and directly affect map granularity and weight behavior.

assumptions (4)

ad hoc to paper Cosine similarity between a pixel's feature vector and its disparity-shifted version measures stereo matching confidence.
Invoked in Section II.A, Eq. (1); no derivation or calibration is provided, and as written it omits the right image feature vector.
domain assumption The iterated extended Kalman filter in ReAqROVIO is consistent, so pose uncertainty can be neglected in the map.
Section II.C states this assumption, conditional on no sudden environmental changes. It is used to justify ignoring state covariance in the confidence representation.
ad hoc to paper Replacing Voxblox's accumulating weight update with a simple average preserves map quality.
Section II.C, Eq. (12); introduced without derivation, ablation, or analysis of statistical consequences for the TSDF estimate.
domain assumption The simplified RGB-D sensor weight model from Voxblox transfers to underwater stereo cameras.
Section II.B relies on [8] and [12] for the z^{-2} weighting; no underwater-specific noise model is proposed or validated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Uncertainty Aware Mapping for Vision-Based Underwater Robots." pith.science (2026). https://pith.science/paper/R5DIQZQL

@misc{pith2026250710991,
  author       = {Pith},
  title        = {Pith review of: Uncertainty Aware Mapping for Vision-Based Underwater Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5DIQZQL}},
  note         = {Machine review of arXiv:2507.10991}
}

read the original abstract

Vision-based underwater robots can be useful in inspecting and exploring confined spaces where traditional sensors and preplanned paths cannot be followed. Sensor noise and situational change can cause significant uncertainty in environmental representation. Thus, this paper explores how to represent mapping inconsistency in vision-based sensing and incorporate depth estimation confidence into the mapping framework. The scene depth and the confidence are estimated using the RAFT-Stereo model and are integrated into a voxel-based mapping framework, Voxblox. Improvements in the existing Voxblox weight calculation and update mechanism are also proposed. Finally, a qualitative analysis of the proposed method is performed in a confined pool and in a pier in the Trondheim fjord. Experiments using an underwater robot demonstrated the change in uncertainty in the visualization.

Figures

Figures reproduced from arXiv: 2507.10991 by the authors.

**Figure 2.** Depth and confidence estimation in RAFT Stereo from disparity field. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Visualization of confidence voxels and voxblox’s surface reconstruction of our field experiment in a pier in the Trondheim fjord. (a) The top view of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visualization of image, depth map and confidence map. (a) and (b) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Mapping in the indoor pool environment at MC-Lab. (a) shows the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[1]

Autonomous inspection of underwater structures,

M. Jacobi, “Autonomous inspection of underwater structures,” Robotics and Autonomous Systems , vol. 67, pp. 80–86, 2015

work page 2015
[2]

Adaptive integral sliding mode control for attitude tracking of underwater robots with large range pitch variations in confined spaces,

X. Wang, Z. Sha, and F. Zhang, “Adaptive integral sliding mode control for attitude tracking of underwater robots with large range pitch variations in confined spaces,” IEEE Robotics and Automation Letters , vol. 10, no. 2, pp. 979–986, 2025

work page 2025
[3]

Safe and efficient path planning under uncertainty via deep collision probability fields,

F. Herrmann, S. Zach, J. Banfi, J. Peters, G. Chalvatzaki, and D. Tateo, “Safe and efficient path planning under uncertainty via deep collision probability fields,” IEEE Robotics and Automation Letters , vol. 9, pp. 9327–9334, 2024

work page 2024
[4]

Modeling surface multipath effects in synthetic aperture sonar,

B. J. Davis, P. T. Gough, and B. R. Hunt, “Modeling surface multipath effects in synthetic aperture sonar,” IEEE Journal of Oceanic Engineer- ing, vol. 34, no. 3, pp. 239–249, 2009

work page 2009
[5]

Online refractive camera model calibration in visual inertial odometry,

M. Singh and K. Alexis, “Online refractive camera model calibration in visual inertial odometry,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2024, pp. 12 609–12 616

work page 2024
[6]

Raft-stereo: Multilevel recurrent field transforms for stereo matching,

L. Lipson, Z. Teed, and J. Deng, “Raft-stereo: Multilevel recurrent field transforms for stereo matching,” in International Conference on 3D Vision (3DV), 2021

work page 2021
[7]

Raft: Recurrent all-pairs field transforms for optical flow,

Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” 2020. [Online]. Available: https://arxiv.org/abs/2003.12039

arXiv 2020
[8]

V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,

H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto, “V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2017

work page 2017

Show all 12 references

[9]

Octomap: A probabilistic, flexible, and compact 3d map representation for robotic systems,

K. M. Wurm, A. Hornung, M. Bennewitz, C. Stachniss, and W. Burgard, “Octomap: A probabilistic, flexible, and compact 3d map representation for robotic systems,” in Proc. of the ICRA 2010 workshop on best practice in 3D perception and modeling for mobile manipulation , vol. 2, ...

2010
[10]

An online self-calibrating refractive camera model with application to underwater odometry,

M. Singh, M. Dharmadhikari, and K. Alexis, “An online self-calibrating refractive camera model with application to underwater odometry,” 2023

2023
[11]

Kinectfusion: Real-time dense surface mapping and tracking,

R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, “Kinectfusion: Real-time dense surface mapping and tracking,” in 2011 10th IEEE international symposium on mixed and augmented reality . Ieee, 2011, p...

2011
[12]

Modeling kinect sensor noise for improved 3d reconstruction and tracking,

C. V . Nguyen, S. Izadi, and D. Lovell, “Modeling kinect sensor noise for improved 3d reconstruction and tracking,” in 2012 second interna- tional conference on 3D imaging, modeling, processing, visualization & transmission. IEEE, 2012, pp. 524–530. Presented at the 2025 IEEE ...

2012

Pith tools

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

[1] [1]

Autonomous inspection of underwater structures,

M. Jacobi, “Autonomous inspection of underwater structures,” Robotics and Autonomous Systems , vol. 67, pp. 80–86, 2015

work page 2015

[2] [2]

Adaptive integral sliding mode control for attitude tracking of underwater robots with large range pitch variations in confined spaces,

X. Wang, Z. Sha, and F. Zhang, “Adaptive integral sliding mode control for attitude tracking of underwater robots with large range pitch variations in confined spaces,” IEEE Robotics and Automation Letters , vol. 10, no. 2, pp. 979–986, 2025

work page 2025

[3] [3]

Safe and efficient path planning under uncertainty via deep collision probability fields,

F. Herrmann, S. Zach, J. Banfi, J. Peters, G. Chalvatzaki, and D. Tateo, “Safe and efficient path planning under uncertainty via deep collision probability fields,” IEEE Robotics and Automation Letters , vol. 9, pp. 9327–9334, 2024

work page 2024

[4] [4]

Modeling surface multipath effects in synthetic aperture sonar,

B. J. Davis, P. T. Gough, and B. R. Hunt, “Modeling surface multipath effects in synthetic aperture sonar,” IEEE Journal of Oceanic Engineer- ing, vol. 34, no. 3, pp. 239–249, 2009

work page 2009

[5] [5]

Online refractive camera model calibration in visual inertial odometry,

M. Singh and K. Alexis, “Online refractive camera model calibration in visual inertial odometry,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2024, pp. 12 609–12 616

work page 2024

[6] [6]

Raft-stereo: Multilevel recurrent field transforms for stereo matching,

L. Lipson, Z. Teed, and J. Deng, “Raft-stereo: Multilevel recurrent field transforms for stereo matching,” in International Conference on 3D Vision (3DV), 2021

work page 2021

[7] [7]

Raft: Recurrent all-pairs field transforms for optical flow,

Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” 2020. [Online]. Available: https://arxiv.org/abs/2003.12039

arXiv 2020

[8] [8]

V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,

H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto, “V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2017

work page 2017