acceptodds
Under review as a conference paper at ICLR 2027

Opti3R: Geometry-Locked Optical Pairing and Reliability-Weighted Cross-View Supervision for Underwater 3D Reconstruction

16 views

Abstract

Attenuation and backscatter change underwater image appearance even when the scene geometry stays fixed. These changes can lead feed-forward models to predict inconsistent depth and shape. To reduce this effect, we propose Opti3R, which trains on Geometry-Locked Optical Pairing (GLOP): paired reference and optically perturbed images with consistent geometry targets. We further use teacher correspondences to supervise cross-view consistency with continuous visibility-confidence-reprojection (VCR) edge weighting, referred to as VCR Graph supervision (VCRG). Our proposed Axial-Channel Geometry Adapter (ACGA) adds identity-initialized multiscale residual blocks to the dense depth and point heads. These blocks take dense features as input and are optimized jointly with the GLOP and VCRG objectives. Raw inference achieves absolute relative depth errors of 0.0838 on SQUID and 0.0496 on FLSea-VI. GLOP gives lower reconstruction error than generic and unpaired augmentation. Opti3R also shows slower growth in geometry drift under held-out optical conditions, and higher VCR edge weights are associated with lower independent geometric disagreement. These results support geometry-locked optical pairing and continuously weighted cross-view supervision for underwater reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, Multi-view Reconstruction, Underwater Vision, Vision Foundation Models, Robust Representation Learning, Cross-view Consistency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4CXsOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a forward underwater 3D reconstruction framework named Opti3R, designed to tackle the challenge of inconsistent depth and shape predictions arising from optical variations such as illumination, attenuation, and backscatter in underwater 3D reconstruction scenarios. Its core contributions are as follows:1. Geometrically Locked Optical Pairing (GLOP): This approach maintains the scene geometry, pixel coordinates, and camera parameters constant while applying optical perturbations solely to the reference image to generate paired samples. This constrains the model to preserve predictive invariance amidst optical changes.2. Visibility-Confidence-Reprojection Graph Supervision (VCRG): Leveraging the correspondence relationships from a teacher model, VCRG integrates visibility, endpoint confidence, and reprojection consistency to assign continuous reliability weights to cross-view supervised edges, thereby preventing the loss of valuable information through hard threshold masking.3. Axis Channel Geometry Adapter (ACGA): ACGA incorporates multi-scale residual blocks, initialized with identity mappings, into the dense depth and point cloud prediction heads. These blocks undergo joint optimization with the objectives of GLOP and VCRG.4. Validation: Comprehensive evaluations were performed on several publicly accessible underwater datasets, including SQUID, FLSea-VI, Sea-thru, FLSea-Stereo, Water3D, and SeaThru-NeRF. The results demonstrate that our framework surpasses existing baselines (e.g., Wat3R, VGGT) in monocular depth estimation, multi-view reconstruction, and relative pose estimation tasks.

Critical strengths
  1. By using the concept of "geometric locking" to decouple optical changes from scene geometry, the core pain point of geometric drift caused by water quality changes in the forward model of underwater images can be directly addressed. The theoretical motivation is sufficient and significant robustness improvement has been demonstrated in experiments.

2.Compared to the binary mask (hard filtering) used in traditional underwater reconstruction, VCRG introduces a continuous three factor weight that balances visibility, confidence, and reprojection consistency. This not only eliminates invalid matches but also preserves valuable mild evidence in low contrast areas, significantly improving the efficiency of cross view supervision.

Critical weaknesses

1.When processing real underwater images, GLOP relies on the optical path variable $l (x) $derived from decoupled teacher depth to artificially apply perturbations. This "approximate re rendering" rather than the pairing of real multi light and multi physical environment acquisition may introduce deviations that do not fully match the degradation of real physical water bodies.

2.Although the article mentions that raw Opti3R only adds a small amount of latency (8.54%) compared to Wat3R, the introduction of EMA teacher network, multi-level axis geometry adapter (ACGA), and complex graph weight calculation makes the specific memory usage and inference frame rate analysis insufficient when facing large-scale streaming reconstruction or resource limited devices.

  1. The overall objective loss function (equation 8) contains multiple weighting coefficients (such as $\ lambda $u=1.0, \ lambda $i=1.0, \ lambda $g=0.5, \ lambda $o=0.1 $, etc.), and the paper did not conduct ablation experiments or quantitative sensitivity analysis on the selection of these hyperparameters and their sensitivity to different water turbidity scenarios.

4.Although excellent performance has been achieved on datasets such as SQUID and FLSea, the teacher correspondence estimation (VCRG) of the model may fail under degraded extreme boundary conditions such as extreme turbidity, strong suspended particles, or complete loss of features caused by multiple scattering. There is a lack of boundary exploration on Failure Cases in the article.

5.The geometric backbone network and camera head of Opti3R are set to a frozen state during training, only adapted through lightweight optical heads and ACGA. Although this maintains the general geometric prior of the pre trained model, it may also limit the model's ability to fully adapt to certain specific complex underwater optical distortions.

Questions for authors
  1. Regarding real data perturbation: In the optical perturbation formula (Equation 2) of actual underwater images, is using the teacher's depth $d (x) $to calculate the optical path sensitive to depth estimation errors? If there is a significant deviation in the depth of the teacher, will the disturbance introduce negative feedback?

  2. Regarding hyperparameter design: Are the weights of each item in the loss function (such as $\ lambda. g=0.5, \ lambda. o=0.1 $) determined through grid search? Do these weights need to be re tuned when facing datasets with a wide range of water quality?

2 comments

Thank you for giving us time to follow through on the tests we outlined. We have added the optical-path, parameter, stress, capacity, and view-scaling results to Supplementary Appendix H. They also helped us make the Discussion more precise about where Opti3R works well and where the evidence becomes thin.

W1 — Real-image optical perturbation. The real-image branch changes appearance using an optical path computed from detached teacher depth. We describe it in Sec. 3.2 as a geometry-preserving training intervention, with the teacher-depth error affecting perturbation strength. We tested that dependence by adding relative depth noise only to the path: Water3D error is 0.332 with the original path, 0.335 at 10% noise, 0.340 at 20%, 0.356 at 40%, and 0.367 with a constant path (Supplementary Table S16). We also checked acquired SQUID images by visibility and both FLSea-VI evaluation scenes; Opti3R improves on Wat3R in each reported group (Table S11). A geometry-calibrated capture of one scene in several real water conditions would extend this test. Sea-Thru and SeaThru-NeRF show why direct attenuation and scattering deserve that kind of controlled acquisition.

W2 — Memory, speed, and scaling. The EMA teacher and correspondence graph are used in training; raw inference runs one student predictor. On a 24-GiB RTX 4090, Opti3R inference takes 0.89 s/17.5 GiB for 8 views, 1.86 s/20.8 GiB for 16, and 3.94 s/23.1 GiB for 32. The corresponding training steps take 2.24 s/22.4 GiB, 4.56 s/23.0 GiB, and 9.37 s/23.8 GiB (Supplementary Table S19). Supplementary Table S7 retains the matched 8-view Wat3R comparison. These are timings for a set of views; sustained streaming FPS would also include input, output, and persistent-state costs. VGGT likewise makes view-count scaling an important part of evaluating feed-forward geometry systems.

W3 — Loss weights and water conditions. We selected the coefficients through a one-factor sweep on held-out DTU validation scans, then kept (λu,λi,λg,λo,τ)=(1,1,0.5,0.1,1)(\lambda_u,\lambda_i,\lambda_g,\lambda_o,\tau)=(1,1,0.5,0.1,1) fixed across the reported underwater evaluations. Supplementary Table S15 gives low/default/high values for every coefficient and τ\tau on overall and severe-water Water3D. For example, λg=0.25/0.5/1\lambda_g=0.25/0.5/1 gives overall errors 0.339/0.332/0.336 and severe-water errors 0.401/0.389/0.396. The default is best among the tested choices in each one-factor comparison. This establishes local sensitivity around the selected setting; a joint grid would answer a different question about interactions.

W4 — Extreme turbidity and feature loss. We have added a controlled stress test that increases attenuation and backscatter while reducing feature visibility on the same scenes. From mild to extreme severity, valid VCR edges fall from 0.83 to 0.21, SQUID relative depth error rises from 0.084 to 0.187, and Water3D error rises from 0.332 to 0.458 (Supplementary Table S17). The revised Discussion names this loss of correspondence evidence as a failure boundary. Real suspended-particle and multiple-scattering sequences with independent geometry would be a valuable next test of that boundary.

W5 — Frozen-backbone capacity. Secs. 3.1 and 4.1 now explain that optical modulation changes the input to the frozen camera head. We also trained two matched-budget unfreezing controls (Supplementary Table S18). The frozen configuration has 14.1M trainable parameters and gives 0.332 Water3D error; unfreezing the camera head gives 15.4M and 0.330; additionally unfreezing the last backbone block gives 27.6M and 0.327. Extreme-condition depth error moves from 0.187 to 0.184 to 0.182. More trainable capacity helps modestly, with a clear parameter cost. AdaptFormer offers a useful general precedent for adapting a fixed vision transformer; our control tests that trade-off for underwater geometry.

Q1 — Teacher-depth error and feedback. The teacher depth used for the optical path is detached during each student update. This blocks a direct gradient path through the perturbation, while the EMA teacher can still carry bias across updates. We isolated path error by changing only the path variable and keeping the teacher geometry target fixed. The measured 10%, 20%, and 40% noise results are 0.335, 0.340, and 0.356 Water3D error, compared with 0.332 for the original path (Supplementary Table S16). The trend shows graceful degradation at smaller errors and a meaningful cost at larger ones.

Q2 — Grid search and retuning. We used a compact one-factor sweep, not a full joint grid search, and used one coefficient setting across the underwater datasets. Table S15 reports the tested values and severe-water sensitivity; for every listed parameter, the chosen default is the best of its low/default/high settings. We have added the selection protocol to Supplementary Appendix H and the final coefficients remain in Sec. 3.5. The sweep supports a fixed choice over the evaluated conditions without implying that one setting is optimal for every water body.

We appreciate your push to test costs and failure modes alongside accuracy. We hope these measurements make the revised paper substantially more useful, and we would be grateful if you would reconsider your assessment.

Reviewer_E3aTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Opti3R to improve the geometric stability of feed-forward underwater 3D reconstruction under changing optical conditions. The method has three components. Geometry-Locked Optical Pairing (GLOP) constructs optically different image pairs while keeping scene geometry and cameras fixed, and applies teacher-student geometric consistency supervision. Visibility-Confidence-Reprojection Graph supervision (VCRG) assigns continuous weights to cross-view correspondences using visibility, teacher confidence, and student reprojection residuals. The Axial-Channel Geometry Adapter (ACGA) adapts dense prediction features through identity-initialized multiscale residual blocks. Experiments evaluate depth estimation, 3D reconstruction, and relative pose on several underwater datasets. The authors report improvements over Wat3R and analyze robustness to optical distribution shifts and correspondence reliability.

Critical strengths
  1. The paper addresses the discrepancy between changes in underwater image appearance and changes in actual scene geometry. GLOP provides a clear training signal by varying optical conditions while holding geometry and cameras fixed, directly targeting the stated problem.

  2. The evaluation covers monocular depth, multi-view depth, Water3D reconstruction, and SeaThru-NeRF relative pose. In Tables 1-3, Opti3R outperforms its main baseline, Wat3R, under the reported protocols. Sections 4.3-4.4 analyze the gains through controlled optical perturbations and correspondence-weight analysis.

Critical weaknesses
  1. The stability of the incremental component gains remains unclear. Section 4.1 (p. 7) states that the main component comparisons use a fixed seed. The five-run training-budget study addresses training duration, but does not establish the variability of the individual gains from GLOP, VCRG, and ACGA. Repeated runs of the central component comparisons would clarify whether the reported improvements are consistent across training seeds.

  2. The mechanism analyses leave specific uncertainties about geometry stability and correspondence reliability. Equation (4) (p. 4) fits a scale and shift to student depth, while point consistency also normalizes scale. The scale-shift-aligned depth objective and depth benchmark metrics remove scale and offset differences from these comparisons. Reporting the fitted alignment parameters under optical perturbations would clarify this variation alongside the existing geometric accuracy results. Although the full model includes supervised and EMA-teacher camera losses, the controlled depth-drift analysis does not directly establish camera consistency under optical changes. In addition, Equation (5) (p. 5) downweights correspondences using the current student's reprojection residual, which may suppress geometrically valid matches that the student currently predicts poorly. Figure 3 and Section 4.4 (p. 9) show a calibration trend on two held-out DTU scenes, but leave its transfer to real underwater correspondence errors uncertain.

Questions for authors
  1. Do the central incremental gains from GLOP, VCRG, and ACGA persist across repeated runs under the same training budget? Could the authors report the mean and variability of the key component comparisons across training seeds?

  2. Under controlled optical perturbations, could the authors report the fitted depth-alignment scale and shift, shape consistency under a stated common gauge, and consistency of relative camera rotation and translation direction? These diagnostics would clarify the roles of scale variation, additive depth bias, and changes in reconstructed shape.

  3. Under matched correspondence construction and validity criteria, could the authors compare uniform weights over valid edges, teacher-confidence weighting, a clearly defined teacher-only residual weighting, and the full VCR weighting, and report sensitivity to τ\tau?

  4. What geometry estimates or independent reference define the "independent geometric disagreement" in Figure 3, and which quantities, if any, are shared with correspondence construction or the VCR weights? Could the calibration analysis also be evaluated on a real underwater subset where an independent geometric reference is available?

1 comment

Thank you for giving us time to finish the measurements we discussed. Your questions exposed an important gap between the original training-budget repetitions and the stability of the actual component gains. We have filled that gap and added the geometry and correspondence diagnostics to the revised paper.

W1 — Stability of incremental gains. We have trained GLOP, GLOP+VCRG, and GLOP+VCRG+ACGA with paired seeds {0,1,2,3,4}\{0,1,2,3,4\}, all at 19,200 updates. Their SQUID relative errors are 0.0965 ± 0.0010, 0.0888 ± 0.0009, and 0.0838 ± 0.0010; FLSea-VI gives 0.0569 ± 0.0006, 0.0527 ± 0.0006, and 0.0496 ± 0.0006; Water3D gives 0.3780 ± 0.0049, 0.3520 ± 0.0049, and 0.3320 ± 0.0060. Each paired seed retains both incremental improvements. Main-paper Table 6 gives the means and sample standard deviations; Supplementary Table S10 shows every seed. The earlier five-run training-duration study remains separate in Appendix E.

W2 — Gauge, cameras, hard matches, and real calibration. We have added measurements that the aligned depth metric could not provide. Supplementary Table S12 reports fitted depth scale and shift, common-gauge shape, and relative camera changes across optical severity. At held-out severity, the Opti3R/Wat3R pairs are 0.046/0.124 for ∣aD−1∣|a_D-1|, 0.021/0.052 for normalized ∣bD∣|b_D|, 0.069/0.137 for shape, 1.63°/2.81° for rotation, and 3.10°/4.90° for translation direction. We also compared four VCR weighting rules on identical candidate-edge support. Full VCR has 0.332 Water3D error and recovers 0.610 of initially difficult but independently valid edges; teacher-only residual weighting gives 0.348 and 0.480 (Supplementary Table S13). On a real FLSea-Stereo subset, weight and independently referenced disagreement have edge-level Spearman correlation −0.63-0.63 (Supplementary Table S14). These results are summarized in Secs. 4.3–4.4.

Q1 — Repeated component results. Yes. The paired-seed study uses the same initialization, data, model-selection rule, and 19,200-update budget for all three variants. The Water3D means move from 0.3780 ± 0.0049 to 0.3520 ± 0.0049 to 0.3320 ± 0.0060 (Table 6); Supplementary Table S10 contains the individual values. This is the component experiment you asked for, distinct from the training-duration study.

Q2 — Scale, shift, shape, and relative cameras. For every reference–perturbed pair, we record the fitted (aD,bD)(a_D,b_D) and divide ∣bD∣|b_D| by the reference median depth. We fix the reference median depth and camera frame once for the shape comparison, then measure relative rotation and translation-direction angles. Supplementary Table S12 gives all four severity levels. At the held-out level, Opti3R's scale/shift/shape/rotation/direction values are 0.046/0.021/0.069/1.63°/3.10°, versus Wat3R's 0.124/0.052/0.137/2.81°/4.90°. This directly tests the variation that scale–shift-aligned depth error hides.

Q3 — Four weights and sensitivity to τ\tau. All four controls use the same candidate edges, finite-geometry screen, bidirectional visibility test, and loss normalization. Uniform, teacher-confidence, teacher-only residual, and full VCR weighting yield Water3D errors of 0.366, 0.354, 0.348, and 0.332 (Supplementary Table S13). The teacher-only rule uses rt=∥ut−x^∥2r^t=\|u^t-\widehat{x}\|_2; full VCR uses the detached student residual rsr^s. The fraction of independently valid difficult edges recovered by training is 0.410/0.460/0.480/0.610 across the same four rules. At τ=0.5,1,2\tau=0.5,1,2 pixels, Water3D error is 0.341/0.332/0.337 (Supplementary Table S15).

Q4 — Source and independence of geometric disagreement. The teacher supplies candidate correspondences and VCR weights. For each selected edge, we compute disagreement from DTU reference geometry or calibrated FLSea-Stereo disparity and cameras in one frame, normalized by the scene's median positive reference depth. This is independent of teacher-predicted geometry conditional on the teacher-selected edges. We now state the construction and its sampling qualification in Sec. 4.4 and Supplementary Appendix H. On FLSea-Stereo, disagreement falls from 0.106 in the lowest weight bin to 0.034 in the highest; the edge-level Spearman correlation is −0.63-0.63 (Table S14). The real-data result is for this calibrated subset, and we would welcome a broader collection of water conditions in future work.

We appreciate how specific your requests were: they made it possible to test the proposed mechanism rather than rely on aligned accuracy alone. We hope these additions give you the confidence to reconsider your score.

Reviewer_V3doOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Opti3R adapts a frozen feed-forward geometry model to underwater images using geometry-locked optical pairs, an EMA teacher–student consistency objective, continuously weighted cross-view supervision, and residual adapters in the depth and point heads. The evaluation covers depth, reconstruction and pose across several underwater datasets and reports consistent gains over Wat3R.

Critical strengths
  1. The geometry-preserving optical perturbation is well motivated: changing underwater appearance should not change the scene geometry. The shuffled-pair control is particularly useful for separating the benefit of scene correspondence from that of the perturbation distribution alone.

  2. Improvements over Wat3R are consistent across several tasks and datasets. The care taken to specify gradient flow and correspondence construction also helps make the approach reproducible.

Critical weaknesses
  1. The distinct contribution relative to existing underwater teacher–student methods is not yet sufficiently clear. GLOP combines mean-teacher consistency with physically motivated augmentation, VCRG uses confidence and reprojection weighting, and ACGA adds residual adapters. These are reasonable choices, but the paper needs a sharper comparison with Wat3R/WAT3R and Water-VGGT to establish what new capability or finding the combination contributes.

  2. The comparisons do not adequately isolate the benefit of the proposed components. The main multi-view, reconstruction and pose results mainly compare against Wat3R, while the key component ablations are deferred to the supplement. Training on FLSea also complicates comparisons against zero-shot models evaluated within the same dataset family. The decisive control is the same backbone, initialization, training data and budget with standard augmentation and supervision. Single-seed main results further limit confidence in the size of the gains. These concerns leave me marginally below acceptance.

Questions for authors
  1. Which frozen backbone and initialization are used? Were Wat3R results re-run with identical resolution, alignment and view selection? Please identify the matched-backbone/data/budget control and summarize its results alongside the key component ablations.

  2. How are the FLSea training and evaluation scenes separated, and what evidence isolates improvement under genuinely different underwater optical conditions from extrapolation within the training renderer?

  3. Does the per-image scale/shift alignment in GLOP permit optical perturbations to change the predicted depth scale? Also, why is the geometric disagreement used to validate VCR weights independent of the teacher geometry used to construct them?

3 comments

Thank you for giving us time to complete the checks we said we would make. We now have the paired-seed component results and direct measurements of how depth, shape, and cameras change under optical perturbation. We have added the main results to the paper and the detailed values to Supplementary Appendix H.

W1 — Distinct contribution and comparisons. We have sharpened the claim in the Related Work and Method: the key training relation is between optically different views of the same geometry. The common-protocol comparison in main-paper Table 2 evaluates Wat3R, WAT3R, Water-VGGT, and Opti3R with the same preprocessing and raw-inference rule. On Water3D, their overall distances are 0.411, 0.371, 0.358, and 0.332, respectively. That comparison establishes performance under a common evaluation protocol. The matched-training test in Table 5 isolates GLOP: with the same Wat3R initialization, data, 19,200 updates, and inference, generic augmentation gives 0.394, unpaired optical perturbation 0.386, and GLOP 0.377. Shuffling the pairs in the full model raises the Water3D result from 0.332 to 0.391 (Supplementary Table S4). The new held-out optical diagnostic also measures common-gauge shape difference: 0.069 for Opti3R versus 0.137 for Wat3R (Supplementary Table S12). Together, these tests show what the pairing contributes and what stability it is associated with.

W2 — Component attribution, FLSea training, and seed uncertainty. We have now repeated GLOP, GLOP+VCRG, and the full model with paired seeds {0,1,2,3,4}\{0,1,2,3,4\} at the same 19,200-update budget. On Water3D, the means ± sample standard deviations are 0.3780 ± 0.0049, 0.3520 ± 0.0049, and 0.3320 ± 0.0060; each paired seed retains the incremental ordering. Main-paper Table 6 reports the three-benchmark summary, and Supplementary Table S10 lists every run. Table 5 and Supplementary Table S4 remain the matched-training controls; Table 2 is a separate checkpoint evaluation. Sec. 4.1 now states this distinction and the FLSea scene partition explicitly. Because FLSea Canyons data are used for training, we do not attribute a cross-method difference to a single component; the matched controls support that attribution.

Q1 — Backbone, initialization, and Wat3R reevaluation. The predictor is VGGT based and starts from the released Wat3R checkpoint. The geometry backbone and camera-head weights are frozen; the optical head, dense depth and point heads, and ACGA blocks are trained. Optical modulation reaches the camera-head input, so the camera predictions can change even with frozen camera-head weights (Sec. 4.1; Supplementary Appendix A). Yes, we reevaluated Wat3R for Table 2 under the common 518-pixel preprocessing, valid masks, depth alignment, view selection, and raw-inference rule. Its row is specific to that evaluation and should be compared within Table 2. The matched-budget GLOP control is Table 5, the detailed ablation is Supplementary Table S4, and the paired-seed summary is Table 6.

Q2 — FLSea separation and different optical conditions. Training uses FLSea Canyons sequences. FLSea-VI evaluation uses Coral Table Loop and U-Canyon, and U-Canyon is excluded from training, teacher-target generation, validation, model selection, and perturbation construction. Scenes are partitioned before optical variants are generated (Sec. 4.1). We have also added a real-image breakdown: across high-, medium-, and low-visibility SQUID strata, Opti3R relative depth error is 0.075, 0.083, and 0.095, compared with Wat3R's 0.094, 0.106, and 0.124. Both FLSea-VI evaluation scenes improve as well (Supplementary Table S11). The [0.8,1][0.8,1] attenuation range remains a separate controlled shift test; we interpret the acquired-image results separately.

Q3 — Depth alignment and the VCR reference. The scale–shift-aligned GLOP depth term by itself cannot reveal a scale or offset change. We have therefore measured the fitted transform and fixed a common reference gauge for shape and camera comparisons. At held-out severity, Opti3R versus Wat3R gives ∣aD−1∣|a_D-1| of 0.046 versus 0.124, normalized ∣bD∣|b_D| of 0.021 versus 0.052, shape difference of 0.069 versus 0.137, relative rotation change of 1.63° versus 2.81°, and translation-direction change of 3.10° versus 4.90° (Sec. 4.3; Supplementary Table S12). For VCR calibration, teacher predictions select edges and determine weights; DTU reference geometry or calibrated FLSea-Stereo disparity and cameras determine the disagreement. The reference error is independent of teacher-predicted geometry for each selected edge, while the set of sampled edges remains teacher-selected (Sec. 4.4; Supplementary Table S14).

Your questions led us to separate evaluation, component attribution, and geometry stability much more clearly. We hope the new evidence resolves the concerns behind your assessment, and we would be grateful if you would consider updating your score.

Thanks for the reply. I still have a few questions.

  1. With 40% noise in the teacher depth, the Water3D error increases from 0.332 to 0.356. Does this mean that GLOP is quite sensitive to teacher-depth errors?
  2. Since the teacher selects the candidate edges in the FLSea-Stereo analysis, could this selection affect the reported correlation between VCR weights and geometric disagreement?
  3. The one-factor sweep shows local sensitivity around the chosen setting. Is it enough to support robustness across different water conditions, or does it only show stability near the selected hyperparameters? Thanks again for the clarification.

Thank you for these questions. They give us a chance to explain the results more clearly.

  1. Yes, errors in teacher depth do affect GLOP. In Table S16, Water3D error rises from 0.332 to 0.335 with 10% added noise and to 0.340 with 20% noise. At 40%, it reaches 0.356, about 7.2% higher than the original result. The effect is small at the lower noise levels and more noticeable at 40%. In this test, we changed only the depth used to make the optical perturbation. The teacher’s geometry target stayed the same. If the teacher is also consistently wrong about the geometry, especially in very poor visibility, the effect could be larger.
  2. Yes, the teacher’s choice of edges could affect the correlation. The −0.63 result on FLSea-Stereo applies only to edges the teacher selected. We used separate stereo reference data to measure disagreement, so we were not checking the teacher against its own predictions. Still, if the teacher mostly picks easy matches, the correlation could be different on the matches it missed. Table S13 compares four ways of weighting the same selected edges and shows that full VCR works best on that set.
  3. The one-factor sweep in Table S15 mainly shows that the result is fairly stable when we change one weight near the chosen setting. We also tested the same weights under a controlled severe-water condition, where the chosen values again did best among those we tried. We used one set of weights across the reported underwater datasets. This gives us some confidence for the conditions we tested. Thank you for staying with us through the discussion. We also appreciate that you have not raised a concern about the novelty of the idea in this exchange. Your questions point to extra tests that would make the evidence stronger. We know the discussion time is short, but we hope these answers and the results already in the supplement ease some of your concerns. If they do, we would be very grateful if you could reconsider your score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.