acceptodds
Under review as a conference paper at ICLR 2027

Which output absorbs the conflict? Gauge-aware attribution for feed-forward 3D reconstruction

Abstract

Feed-forward 3D reconstruction models emit depth and cameras from a single pass, and we ask which of those outputs absorbs the disagreement when a scene's geometry conflicts with what the image looks like. The question is not well posed as it stands: for a model defined only up to scale a global depth rescale and a baseline rescale are observationally equivalent, so any "depth did x%, cameras did y%" split rests on a gauge convention the reader cannot see, and we fix one and compare only invariants. Nor is the natural statistic safe: the projection coefficient such decompositions invite is the product of directional alignment and magnitude, , rewarding a component for being large rather than for being right. Our instrument is a ray-preserving warp: every scene point moves along its ray through the reference camera, so that image is byte-identical at every dose while the other views change. The monocular cue is therefore held fixed by construction rather than by assumption, the re-rendered views stay geometrically self-consistent with a different relief instead of being degraded, and the conflict has a dose. Across 224 measurements, 104 (46%) score where explained variance is negative, worse than predicting no change, the worst at against ; an earlier stage of this analysis relied on such a coefficient, and we correct it here. Under explained variance VGGT's outputs jointly reproduce 51–94% of the conflict across four doses, declining monotonically in , though the registered form was not supported, while the relief it reports stays nearly inert (parallax-reliance slope , against 0.719 for RAFT triangulation with ground-truth poses at the same model-input resolution). DUSt3R and Fast3R reproduce none of it, in different ways, and the model whose outputs do reproduce it is the one with an explicit camera head, though it also differs in attention structure, so we state the architectural attribution as supported rather than established. Our identifiability claim is scoped to pointmap scale: a pre-committed rule declined to extend it to a joint disparity gauge over , and a normaliser-frozen control confirmed the decline was not an artifact.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, attribution, identifiability, gauge invariance, cue conflict, feed-forward 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hjHWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper measures the behavior of several feed-forward 3D reconstruction models when the appearance of the input is inconsistent with multi-view correspondences, e.g., a painting. It uses a concept from statistics called "explained variance" to measure the error when perturbing the depth without changing the appearance of the image. But this paper is very hard to read and seems to explain nothing significant.

Critical strengths

No obvious strength.

Critical weaknesses

First, this paper is very hard to follow

  • It contains unexplained symbols in the abstract, e.g., M, theta and R2R^2 that is only explained in the paper.
  • It is very hard to understand what does statistics and natural statistic mean in the context of a feed-forward model.
  • A lot of words are hard to understand, e.g., "gauge" I would call it normalization. "cue-conflict instrument" I would paraphrase it. "rather than degraded" how to degrade? what type of degrade? "pre-register" what is that.
  • The contribution is also hard to understand.
  • The Figure 3 where the "final form" is defined is not in the main text at all.
  • It does seem no previous work handled this attribution problem. But I am not convinced it is an important problem. It is natural to measure robustness. But for the attribution of conflicts, we never have the case of perturbing a real RGB image in real-life, and for painting, previous works already have robustness analysis, including a lot of other challenges. And this paper neither explained the attribution clearly to me nor explained its importance.
Questions for authors

Please review the rewriting of your AI tool to make it easier to understand.

1 comment
AuthorsOpenReview ↗

Reply to Reviewer hjHW

Thank you for the review. The revision is rewritten for readability; the main changes are below.

Readability.

  • The abstract has no symbols.
  • "Gauge" is introduced once, in plain words: these models predict a scene only up to an overall size, so depth and camera distance can be scaled together without changing the images, and comparisons must not depend on that choice.
  • "Pre-registered" is now explained where it is first used: we wrote down our predictions and pass/fail rules before running the experiment. The registration history itself is in Appendix H.
  • "Cue-conflict instrument" and "rather than degraded" are replaced by a direct description of the intervention.
  • Old Figure 3, which defines the measure, is now Figure 2 in the main text, and a new opening figure (Figure 1) shows the intervention and the main result together.
  • Section 4 opens with the four terms the analysis relies on (the score, agreement with the images, the baseline ratio, the camera-scale pair) and says how they differ.

What is measured. The paper does not measure the error from perturbing depth.

  • The intervention: we change the 3D shape of the scene seen by the other cameras, while the main image stays exactly the same.
  • The question: which of the model's outputs (depth, camera position, camera orientation, focal length) change so as to explain the changed images.
  • The measure: "explained variance" scores how closely the image change implied by a set of outputs matches the change the warp imposes. One is a perfect match, zero means no better than predicting no change, and negative means worse.

Why it matters. These models return depth and cameras together, and downstream users take one or the other: a depth map, or a camera trajectory. When the views imply one shape and the main image another, it matters which of these goes wrong. The main finding:

  • VGGT's depth stays close to its prediction for the original scene.
  • It explains the changed images by changing two camera outputs together: the baseline-to-depth ratio and the focal length.
  • Its baseline-to-depth ratio ends up between 0.52× and 2.16× the true value, while its outputs still agree with the images.

So agreement with the images does not certify the 3D result.

On real images. Real scenes are not warped in practice. The warp is a controlled probe, in the same spirit as cue-conflict stimuli used to test image classifiers: real data cannot tell us which output should have moved, because real data do not let us set the conflict and know its size. Painting-style robustness tests measure how much accuracy drops. We ask a different question: which output changes, and whether that change is consistent with the images.

If any passage of the revision remains unclear, we will revise it.

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

This paper investigates how different outputs of feed-forward 3D reconstruction models respond when the reference image is held fixed while the scene geometry underlying the other views is changed. The authors interpret this intervention as a conflict between monocular appearance priors and multi-view geometric evidence. They apply a radial scene deformation centered on the reference camera and re-render the source views to construct inputs with adjustable intervention strength. They then account for the scale ambiguity between depth and camera translation, and analyze output component swaps and projection changes to examine the roles of depth shape and camera-related components.

The main finding is that VGGT’s relative depth shape barely follows the imposed deformation, while its joint outputs reproduce a substantial portion of the target projection changes. DUSt3R and Fast3R do not exhibit the same adaptation. The paper also shows that a projection coefficient can misleadingly credit responses that are excessively large but poorly aligned with the target change. The authors therefore use R² relative to a no-change baseline and report corrections to their earlier conclusions.

Critical strengths

S1. The paper poses a focused diagnostic question at the output level.

Rather than reporting only overall reconstruction error, the paper asks whether depth and camera outputs adapt differently to changes in the input. Holding the reference image fixed provides a concrete intervention for investigating this question. The differences between VGGT and the other two models further motivate examining jointly predicted outputs separately.

S2. The attribution methodology is subjected to careful self-tests and corrections.

The authors explicitly address global scale ambiguity, distinguish individual component swaps from joint swaps, and explain why a projection coefficient cannot be directly interpreted as a fraction of contribution. Testing the estimator using synthetic outputs with known answers, and openly reporting overturned interpretations and unsupported predictions, strengthens the methodological discussion.

Critical weaknesses

W1. Occlusion changes and reprojection artifacts are not sufficiently controlled, which affects the interpretation of the central experiment.

Radial deformation changes occlusion relationships in the source views. Using each source view’s own RGB-D data, depth buffering, and visibility filtering is necessary and reasonable. However, single-layer RGB-D cannot recover surfaces that become newly visible after deformation but were not originally recorded. Background-favoring neighborhood filling only approximates these regions and may introduce texture propagation and boundary artifacts.

The code includes a control that preserves the original geometry while reproducing the hole pattern and applying hole filling. However, I could not find complete corresponding results for all three evaluated models on the primary metrics in the provided materials. The supplementary package also lacks the rendered inputs and hole masks, making it difficult to inspect the severity of these artifacts directly. The visibility implementation requires clarification: do the coverage conditions for combined views ensure valid correspondences for every point–view pair included in the statistics, and does determining visibility from hole-filled depth affect the results?

These issues affect the interpretation that the measured responses are caused by geometric cue conflict. Simply excluding holes from the evaluation region may not be sufficient, because artifacts elsewhere in the input can still influence predictions for other regions and camera parameters.

W2. The restricted viewpoint geometry weakens the experiment’s ability to distinguish competing explanations and limits the interpretation of the observed responses.

The evaluation uses only synthetic driving data, primarily involving forward motion with broadly aligned camera orientations. This is directly relevant to the central experiment: a radial deformation centered on the reference camera may be more readily approximated by changes in camera parameters when viewpoints are nearby and motion directions are restricted. The current results do not establish how strongly VGGT’s combination of limited depth-shape change and camera-side response depends on this viewpoint geometry.

Consequently, even if the joint outputs reproduce the target displacements well under the evaluated conditions, it remains unclear whether this reflects a more general output preference or an explanation that is particularly accessible under specific motion and parallax conditions. The triangulation control with ground-truth camera poses demonstrates that geometric information is recoverable, but it cannot by itself exclude this capacity for compensation when depth and cameras are estimated jointly.

Views captured along an orbit, larger relative viewing angles, and lateral motion would expose the same deformation from different directions and test whether camera adjustments can still substitute for changes in depth shape. Such evaluations could better distinguish competing explanations and establish whether the conclusions depend on forward-motion conditions. Comparisons should also control for, or stratify by, target displacement magnitude, valid coverage, and rendering artifacts, rather than merely increasing input difficulty.

Furthermore, the current results do not rule out structured out-of-distribution failure: the model may interpret unfamiliar scene deformations as familiar shapes observed under different camera parameters. Priors learned from training data are themselves a source of model bias, so behavior under this particular intervention should be distinguished from a generalizable geometric preference. The authors acknowledge that the architectural comparison is not causal, but the discussion of risks to practical systems still requires more direct external validation. The supplementary correlation check on inputs without scene deformation also does not establish the anticipated connection.

The evaluation covers only three models and omits other feed-forward geometric foundation models, such as π3X. However, compared with the viewpoint geometry and interpretation issues, I consider this a secondary limitation in coverage and would not treat the omission of any particular model as an independent reason for rejection.

Questions for authors

Q1. Occlusion and visibility: Please clarify how newly occluded and newly revealed regions are handled. Does the main experiment independently verify visibility for every point–source-view pair? Could the “visible in at least one source view” coverage condition and visibility checks based on hole-filled depth allow point–view pairs without valid visible correspondences to enter the statistics? Could you provide a sensitivity analysis using stricter visibility criteria?

Q2. Holes and rendering artifacts: Could you provide representative full rendered images, hole masks, and hole fractions at each deformation strength? What results does the matched-hole control produce for the evaluated models’ depth responses and joint R²? To what extent do these controls account for the boundary changes and resampling effects accompanying the geometric deformation?

Q3. Viewpoint geometry and experimental discrimination: Do the existing results vary with relative viewing angle, baseline magnitude, or motion direction? Could you stratify the existing data or provide a limited additional experiment involving lateral motion and orbiting views to test whether “camera-side absorption” depends on the particular geometry of forward driving?

Q4. Prior conflict and the out-of-distribution explanation: Holding the reference image fixed ensures that its pixel information is unchanged, but how do you verify that this intervention actually creates a disagreement between the model’s monocular shape preference and the multi-view evidence? What evidence distinguishes this interpretation from out-of-distribution degradation on unnatural scene deformations?

The answers to these questions would directly affect my assessment. Evaluating additional models, such as π3X, is primarily a suggestion for broadening the scope of the study.

2 comments
AuthorsOpenReview ↗

Reply (1/2): visibility, holes and rendering artifacts

Thank you for the careful review. Q1 identified a defect in the submitted evaluation; the revision corrects it.

Q1, visibility. The submitted evaluation scored every source view of a reference point that was visible in at least one view, and it decided visibility using hole-filled depth.

The revision applies a strict rule. A point is scored in a source view only if it is visible in that view, both without the warp and at the dose being scored. Visibility is decided by matching the point's projected depth to that view's rendered depth before any hole filling, within 5%. Newly occluded points are therefore not scored in the views that lose them. A point that lands in a newly revealed region has no recorded surface there, so it fails the test and is not scored.

Effect of the correction (Section 6, Table 3; every changed number in Appendix C):

  • Pairs removed: 21.3–23.5% of scored pairs in the three-view set fail the strict test; 2.3% or fewer in the single-view sets.
  • VGGT: single-view cells change by at most 0.004 in R². Three-view cells rise by 0.10–0.20. All 16 confidence intervals still clear the 0.4 threshold fixed before the experiment (lowest lower bound 0.447).
  • DUSt3R and Fast3R: no interval has its lower bound above 0.4 (six of DUSt3R's intervals are wide enough for their upper bound to pass it). DUSt3R is now negative in all 16 cells.
  • Same pairs at every dose: if every dose is scored on the same point–view pairs, no VGGT cell changes by more than 0.03. Its decline with dose remains in all four view sets.

Q2, holes and artifacts.

  • Examples. Appendix E (Figure 10, Table 7) shows representative rendered inputs at every dose, the hole masks before filling, and the matched control inputs. The smearing introduced by hole filling is visible in these examples.
  • Hole size. Holes cover 1.5–9.8% of a source view on average and up to 15% at the 90th percentile. They grow with the distance of the view from the reference and are largest at both extreme doses.

The matched-hole control removes exactly the pixels the dose removes and fills them the same way, but leaves the scene geometry unchanged. We scored it for all three models in all 16 cells, on the same strict support as the warp, against the change the warp imposes. Its effect on all outputs is measured jointly (depth, cameras and the projection they imply), not only inside the holes. This tests whether artifacts affect predictions outside the holes, including the camera parameters.

R² of artifact-only responseoutput change relative to the warp's (median)
VGGT−0.03 to +0.0712% (max 18%)
DUSt3R−3.12 to −0.0747%
Fast3R−1.67 to −0.1759%

For VGGT, holes and their filling alone do not reproduce the response. Its artifact-only depth-shape response stays below 0.001 in R².

DUSt3R and Fast3R move about half as far under artifacts alone as under the warp, in directions unrelated to the warp. This sensitivity to artifacts means their negative scores cannot be attributed to the conflict specifically. Section 5.4 therefore reads their results as instability under this intervention, not as a different way of handling the conflict.

What remains open (Sections 6 and 7). The resampling of boundaries that accompanies the geometric change cannot be produced without that change. Interactions between holes and the warp are also untested. We state both as limits.

AuthorsOpenReview ↗

Reply (2/2): viewpoint geometry, OOD, and scope

W2/Q3, viewpoint geometry. We could not render or run lateral or orbiting sequences within the discussion period, so the revision makes no claim beyond forward motion. It names lateral and orbiting motion as the main open test (Section 7).

Within the existing data we stratified VGGT's per-frame agreement (Spearman correlation of per-frame explained variance with each covariate):

  • Baseline: agreement tends to rise with the baseline when more distant frames are included, most clearly four frames apart (up to +0.76; interval above zero in all four doses there, 8 of 16 cells overall). One frame apart, there is no trend.
  • Imposed displacement: agreement also rises with displacement, up to +0.71 (interval above zero in 13 of 16 cells).
  • The two are confounded: baseline and displacement are themselves correlated at +0.37 to +0.88 in forward driving, so we cannot separate them (Appendix F, Table 8).
  • Coverage: no positive association with valid coverage.

So within this regime, larger baselines do not make the camera-side explanation less available. This analysis is descriptive, and it does not answer your question about other motion directions.

The paired-swap result bears on this question but does not settle it. The change is carried jointly by the baseline-to-depth ratio and the focal length (Section 5.3). Under near-forward motion, both of these, like a change of relief, displace image points roughly radially. This may be why the compensation is available here; the revision states it as an untested hypothesis (Section 7).

Q4, does the intervention create a conflict, and is this OOD?

Conflict.

  • The reference image is byte-identical at every dose, so any single-image depth inferred from it cannot change.
  • The other views carry the new relief: triangulation with ground-truth poses at the models' input resolution recovers it (parallax slope 0.719, against 0.012 for VGGT).
  • VGGT's depth barely responds. At every dose, its depth error against the warped truth is within 0.01 of the error of its own unwarped depth.

This is the operational sense in which the reference image and the other views disagree.

OOD. Structured out-of-distribution behaviour cannot be excluded: an unfamiliar shape may be read as a familiar shape seen by different cameras, which is consistent with what we observe at the output level. The response is, however, not well described as an indiscriminate breakdown of the outputs:

  • Rotation and direction stay within 0.2° and 1.1° of the truth.
  • The depth shape stays close to its unwarped value.
  • The outputs reproduce the imposed change with explained variance 0.53–0.94.
  • Artifact-only inputs move the outputs a median 12% as far as the warp does.

Whether the outputs respond to the conflict itself or to unfamiliar geometry as such, the data do not decide. The revision therefore describes the result as output behaviour under this intervention, not as an internal preference of the model (Section 7).

The check on unmanipulated frames did not establish the expected relation. It is reported as a null, and we have removed the recommendation about downstream systems from the discussion.

π3X and other models. We did not add models in the discussion period. The limitation is stated in Section 7.

Reviewer_at2YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates how feed-forward 3D reconstruction models (VGGT, Fast3R, and DUSt3R) distribute geometric inconsistencies between depth and camera predictions when monocular priors conflict with multi-view parallax cues. The authors first establish that directly separating depth and camera contributions is ill-posed due to the scale-gauge equivalence (d,t)↦(λd,λt)(d,t)\mapsto(\lambda d,\lambda t), requiring gauge fixing and comparisons based on gauge-invariant quantities. They further demonstrate that the commonly used projection-based attribution coefficient β=Mcos⁡θ\beta=M\cos\theta can produce substantial false positives by rewarding response magnitude, and propose using the uncentered explained variance R2=cos⁡2θ−(M−cos⁡θ)2R^2=\cos^2\theta-(M-\cos\theta)^2, which penalizes magnitude mismatches instead. Experimentally, the authors introduce ray-preserving warping, which moves 3D points along reference-camera rays to introduce controlled multi-view parallax conflicts while keeping the reference image byte-for-byte identical. Experiments on Virtual KITTI 2 show that the depth variation patterns of all three models remain largely insensitive to parallax conflicts, predominantly following monocular appearance priors. However, only VGGT, which has an explicit camera regression head, can jointly absorb and reproduce 51%–94% of the geometric conflicts through adjustments to camera baseline ratios and related parameters. In contrast, Fast3R and DUSt3R, which rely on solver-based or pairwise alignment mechanisms, exhibit unstable camera poses or point clouds (R2<0R^2<0). The authors also report negative results for the joint focal-length–baseline disparity-gauge hypothesis and the monotonicity hypothesis concerning ∣log⁡γ∣|\log\gamma|, strictly following their preregistered evaluation criteria.

Critical strengths
  1. Meaningful and relatively novel research question. The paper investigates how feed-forward 3D reconstruction models distribute conflicts between monocular priors and multi-view geometry across depth and camera predictions. This provides a valuable perspective beyond conventional reconstruction accuracy and robustness evaluations.

  2. Well-designed interventions and rigorous analysis. Ray-preserving warping introduces controlled geometric conflicts while keeping the reference image unchanged. The treatment of gauge ambiguity and the use of uncentered explained variance further improve the validity and interpretability of the analysis.

  3. Insightful findings on model behavior. Comparisons across VGGT, Fast3R, and DUSt3R reveal distinct responses to geometric conflicts. In particular, VGGT maintains relatively stable normalized depth structures while adjusting camera predictions to accommodate parallax changes, providing useful evidence for understanding the interaction between monocular priors and multi-view geometry.

Critical weaknesses
  1. Insufficient experimental validation. The main paper presents limited experimental results, with important analyses relegated to the appendix. Evaluation is restricted to three reconstruction models and a single synthetic driving dataset, limiting the generalizability of the findings. More extensive comparisons, controlled architectural ablations, and deeper analyses of the observed behaviors are needed to strengthen the conclusions.

  2. Limited clarity and organization. The paper devotes considerable space to methodological corrections, unsuccessful preliminary experiments, and preregistration details, which obscures its central contributions. Meanwhile, key experimental results receive insufficient attention in the main text. A clearer structure, more intuitive explanations, and a better balance between methodology and experimental analysis would improve readability and persuasiveness.

Questions for authors

1. How well do the findings generalize to other datasets and reconstruction models? The experiments are limited to three models and a single synthetic driving dataset. Could the authors provide additional evaluations across different scene types, camera motions, or reconstruction architectures?

2. Does a higher explained variance correspond to more accurate 3D reconstruction? VGGT achieves high explained variance in its output changes, but this does not necessarily indicate accurate geometry. Could the authors report depth and camera pose errors to distinguish meaningful geometric adaptation from compensating output changes?

3. What are the practical implications of the observed conflict allocation for downstream applications? The paper suggests that changes in camera estimates under geometric conflicts may affect downstream systems. Could the authors provide concrete examples or quantitative evidence of such effects, particularly in tasks relying on accurate camera poses or 3D geometry?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer at2Y

Thank you for the in-depth review of the paper. Q2 led to the analysis that is now the paper's main result (Sections 5.2–5.3).

Q2, does high explained variance mean accurate 3D? No, and the revision now shows it (Section 5.2, Table 2, Figure 1C–D). Because the intervention fixes the ground truth of every condition (the warped depth, and cameras the warp leaves unchanged), we measured each model's error against it directly. For VGGT:

without the warpacross the four doses
depth-shape error0.03–0.040.13–0.57
baseline/depth ratio, estimated ÷ true0.97–0.990.52–2.16
rotation error≤ 0.1°≤ 0.2°
translation-direction error0.2–0.7°0.3–1.1°

At every dose, VGGT's depth-shape error is within 0.01 of the error of a depth map that did not change at all. What moves is the camera scale: the baseline-to-depth ratio and the focal length.

Paired swaps show how this works (Section 5.3):

  • Baseline-to-depth ratio alone: worse than predicting no change in all 16 cells.
  • Focal length alone: worse in 9 of 16, and reaching 0.4 in only 2.
  • The two together: R² 0.49–0.94, a median 98% of what all outputs achieve.

So the high explained variance comes from a coupled camera-scale change, not from accurate geometry. A one-output-at-a-time attribution would miss it.

DUSt3R and Fast3R also err in camera scale (Table 2). However, their rotation and direction errors are larger, and so is their response to rendering artifacts alone.

W1/Q1, validation and generalization.

  • More of the evidence is now in the main text. The swap results for all three models, the accuracy table, the controls table and the component figure are all there, rather than in the appendix.
  • Internal validity is stronger. The revision adds strict per-view visibility, a matched rendering-artifact control and a fixed-support check for all models and cells (Section 6). In the process we found and corrected an evaluation defect (general comment); the correction reverses no claim (Appendix C).
  • We have not added datasets, motions or models in the discussion period. The revision states the scope (one synthetic driving dataset, forward motion, three public models) wherever the result is stated.
  • Attributing the differences to architecture would require models trained identically except for architecture. We therefore present the model differences as a comparison of whole models, not a causal attribution (Section 5.4).

W2, clarity. The paper has been reorganised around the finding: physical question, intervention, measurement, results, controls. The registration history and the details of the measure's self-test have moved to the appendix (general comment).

Q3, downstream implications. We have no downstream-task experiment. The check we ran on unmanipulated frames was null, so we removed the recommendation about downstream systems.

What the measurements do imply, stated as an implication: under this conflict, VGGT returns accurate camera rotation and direction but a baseline-to-depth ratio between 0.52× and 2.16× its true value, while its depth stays close to its unwarped prediction. A downstream use that takes the predicted baseline-to-depth ratio as it is, without recalibrating or independently correcting the scale, would inherit that error. We have not measured how much this matters for any particular task. Agreement with the input images does not certify the geometry (Section 5.2).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision, including a correction

We thank the reviewers. All three found the submitted paper hard to read: its findings were buried under methodology, registration history, and limitations. The manuscript is revised for readability. Reviewer yBst also raised substantive questions about validity: whether every scored point is visible, whether rendering artifacts alone could drive the response, and whether the result depends on the viewpoint geometry. The visibility question exposed a real defect, which we correct (item 3); the artifact question is addressed by the matched-hole control (item 2); the viewpoint question remains open and is stated as such. We have uploaded a revised PDF (main text within the 10-page limit), and we summarise the changes here; the individual replies refer to this summary.

1. Rewritten around the finding. The abstract has no symbols. The introduction starts from the physical question: when the other views imply a different 3D shape while the main image stays exactly the same, which predicted quantity changes? The paper now opens with a figure that shows the intervention and the main result together (Figure 1), and the figure that defines the measure (old Figure 3) is now in the main text (Figure 2). Terms such as "gauge" are defined in plain words where they first appear. The registration history, the details of the measure's self-test and the chronology of the corrections below are in the appendix, so the main text reads as one account.

2. New analyses, all on stored predictions. No model was re-run.

  • strict per-view visibility (Section 6);
  • the matched-hole control, with all three models across all 16 dose × view-set cells (Section 6, Table 3);
  • depth and camera error against ground truth (Section 5.2, Table 2);
  • paired output swaps (Section 5.3, Figure 4);
  • a fixed evaluation support across doses (Section 6);
  • stratification by baseline, imposed displacement and coverage (Appendix F).

The matched-hole inputs and their predictions were produced with the original experiment, but they were not analysed before submission. Every analysis added after review was run against a decision rule written down before computing it. The scripts and result files for all of them, and a script that recomputes every number in the revision from those files, are in the updated supplementary material.

3. A correction. Reviewer yBst asked whether every scored point–view pair is truly visible. It was not. The submitted evaluation scored every view of a point that was visible in any one view, and it used depth after hole filling. In the three-view set, 21.3–23.5% of the scored pairs fail a strict per-view test. Every number in the revision that depends on which views are scored now uses the strict rule, and Appendix C lists each number that changed. The main changes:

  • VGGT's joint explained variance of the imposed change is 0.53–0.94 (previously 0.51–0.94).
  • Of 224 measurements, 101 (previously 104) give a positive projection coefficient while explained variance is negative.
  • DUSt3R is now negative in 16 of 16 cells (previously 15).

The correction reverses no claim that depends on these numbers. We also corrected a wording error in the old §6.2 (the cosine range 0.00–0.09 described Fast3R's joint output change, not its rotation term), and withdrew the old §6.3 statement that a correspondence-based score cannot see the camera-scale error: the baseline-to-depth ratio is scale-free and constrained by the images (Section 5.2).

4. What is now claimed, and what is no longer claimed.

  • The central result is more specific. VGGT's depth shape stays where it is without the warp, yet its outputs reproduce most of the image change. They do so through two camera outputs acting together:
    • The baseline-to-depth ratio alone makes the prediction worse than predicting no change in every cell.
    • The focal length alone is worse than predicting no change in 9 of 16 cells and reaches an explained variance of 0.4 in only 2.
    • Swapped in together, they reach explained variance 0.49–0.94, a median 98% of the value with all outputs swapped in (Section 5.3).
  • The outputs agree with the images while describing the scene wrongly. VGGT's ratio of baseline to scene depth ranges from 0.52× to 2.16× the true value (Section 5.2).
  • Three claims are narrower. The architecture comparison is a comparison of whole models, not a causal attribution. The recommendation about downstream systems has been removed. The scope to forward motion is stated wherever the result is stated (Section 7).

TL;DR of the revision. The TL;DR field cannot be edited during the discussion, so the one shown above still describes the submitted version. For the revision: when the other views imply a different 3D shape, VGGT's depth barely changes while its baseline-to-depth ratio and focal length shift together to fit the images; measuring this needs scale-free comparison and explained variance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.