acceptodds
Under review as a conference paper at ICLR 2027

Gate3R: Training-Free Prior Gating for Feed-Forward 3D Reconstruction

Abstract

Feed-forward reconstruction models such as DUSt3R and its successors recover geometry from unposed images in an end-to-end manner. Some of these models are prior-conditioned, taking optional geometric priors such as camera intrinsics, relative pose, and depth to further improve reconstruction. In practice, however, such priors are often inaccurate or corrupted, and conditioning on them can produce reconstructions worse than those from images alone. We propose Gate3R, a training-free method to discard unreliable priors at test time and fall back to image-only reconstruction. It combines the residual between the prior and the prediction with the frozen model's attention entropy to decide whether each prior is usable. On two prior-conditioned models, Pow3R and MapAnything, Gate3R outperforms both always keeping and always discarding priors in most single-prior settings and in all multi-prior settings across six datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed-forward 3D reconstruction, attention entropy, training-free

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_GcpkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

Gate3R addresses the reliability of prior-conditioned feed-forward 3D reconstruction under inaccurate geometric priors. While camera intrinsics, relative pose, and depth can improve reconstruction, corrupted priors may degrade performance below image-only inference. The authors propose a training-free test-time prior-gating method that extracts two signals from a single conditioned forward pass: the prior residual and the attention entropy of a preselected head. These signals are standardized using deployment-batch statistics and combined through tail-weighted fusion. The method then discards the corresponding priors for a fixed fraction of view sets with the highest fused scores and reruns inference. Under controlled synthetic corruptions, Gate3R outperforms both always retaining and always discarding priors on eight of nine single-prior evaluation axes and all 27 multi-prior metric cells, averaged across test datasets.

Critical strengths

1. Practical relevance.
The paper treats whether to trust an input prior as an explicit decision problem, rather than assuming that additional geometric information is always beneficial. This is relevant to reconstruction pipelines that use outputs from other estimators as priors. Gate3R requires no updates to the reconstruction model’s parameters, facilitating integration with existing models.

2. Comprehensive controlled evaluation.
The experiments cover three prior types, five corruption levels, single- and multi-prior settings, and transfer from two to four views. All policies share the same corruption draws across ten random seeds, while attention heads are selected on a held-out dataset. This paired design helps reduce variability in comparisons.

3. Clear average gains over fixed policies.
For two-view MapAnything, Gate3R reduces intrinsics ray-direction error from 2.20° with all priors to 1.79° and achieves a pose AUC@5° of 36.1, compared with 28.4 for retaining all priors and 29.0 for image-only inference. The multi-prior results also show consistent improvements in aggregate.

Critical weaknesses

1. Insufficient Explanation of Attention Entropy. The paper provides limited insights into why attention entropy can indicate conflicts between geometric priors and visual evidence. Although empirical results demonstrate correlations between entropy changes and prior corruption, the underlying mechanism remains unclear. Such changes may simply result from perturbations in prior embeddings rather than genuine geometric inconsistencies. The current visualizations are insufficient to distinguish these possibilities. More interpretable analyses, such as layer-wise attention studies or geometric correspondence visualizations, would strengthen the justification for this approach.

2. Limited Evaluation Across Different Numbers of Views. The experiments are restricted to two-view and four-view reconstruction settings. It remains unclear whether Gate3R can generalize effectively to larger numbers of input views. Increasing the number of views may substantially change attention distributions and the aggregation of prior residuals. Consequently, the selected attention heads and gating strategy may behave differently in larger multi-view settings. Additional experiments with more diverse view counts would better demonstrate the scalability and robustness of the proposed method.

3. Limited Validation Under Real-World Prior Corruptions. The evaluation primarily relies on synthetically corrupted geometric priors, including predefined perturbations to camera intrinsics, poses, and depth. However, real-world prior estimation errors are often more complex and may exhibit spatially varying noise, systematic biases, or correlations across views. It is unclear whether Gate3R can reliably identify such naturally occurring errors. The current corruption protocol may not fully capture the challenges encountered in practical reconstruction pipelines. Experiments using priors estimated by real calibration, visual odometry, or depth estimation methods would strengthen the practical relevance of the approach.

4. Limitations of the Fixed Discard Budget. Gate3R adopts a fixed discard ratio of ρ=0.3\rho=0.3 across different datasets and corruption settings. Although the authors provide budget ablations, their results indicate that the optimal discard ratio varies with the proportion of reliable priors. A fixed budget may unnecessarily discard accurate priors when corruption is rare or retain harmful priors when corruption is widespread. This limitation becomes particularly relevant when the corruption rate is unknown in real-world deployment. An adaptive gating strategy based on estimated prior reliability could potentially improve both reconstruction accuracy and computational efficiency.

5. Ambiguity in Attention Entropy and Head Selection. Attention entropy may reflect intrinsic scene difficulty or visual ambiguity rather than prior corruption alone. The paper does not sufficiently disentangle these factors when interpreting entropy changes as indicators of unreliable priors. Moreover, the attention heads are selected based on their clean-versus-corrupted detection AUC rather than their actual contribution to downstream reconstruction performance. A head with strong corruption detection capability does not necessarily provide the most effective gating decisions, particularly when mildly corrupted priors remain beneficial. Further analysis of these potential confounding factors and alternative head-selection criteria would strengthen the methodological justification.

Questions for authors

All my concerns have been detailed in the weaknesses section.

4 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

We sincerely thank the reviewer for the detailed and constructive comments. We take every comment seriously and hope our response addresses the reviewer's concerns. If any questions remain, we are happy to address them.

Q1: Do entropy changes reflect genuine geometric inconsistencies or merely perturbed prior embeddings? Layer-wise attention studies would strengthen the justification.

The selected head appears to respond to prior–image disagreement. It separates corrupted poses from clean ones of the same rotation angle, whereas removing the prior altogether shifts entropy in the opposite direction or barely at all. Layer-wise, the signal is absent at the injection layer and in Pow3R's pose-blind encoder, growing only in deeper layers where visual evidence meets the prior, with all deployed heads in the network's second half (median head AUC, severe vs. clean, per depth quarter):

Q1Q2Q3Q4Deployed head
MapAny K0.520.520.520.530.59 (Q4)
MapAny P0.530.600.670.700.81 (Q3)
MapAny D0.510.520.520.570.75 (Q3)
Pow3R K0.520.510.510.540.83 (Q3)
Pow3R P0.500.500.500.700.98 (Q4)
Pow3R D0.510.500.510.520.64 (Q4)

Q2: Limited evaluation across different numbers of views.

We evaluate 8 and 16 views on MapAnything with the two-view heads and ρ = 0.3 unchanged; the frozen heads' detection AUC rises with more views, and Gate3R beats both fixed policies on every axis (All / None / Gate3R):

viewsK ray err ↓P AUC@5° ↑D inlier ↑
22.20 / 2.59 / 1.7928.4 / 29.0 / 36.10.393 / 0.375 / 0.435
42.21 / 2.56 / 1.7624.9 / 29.4 / 33.10.301 / 0.324 / 0.355
82.13 / 2.51 / 1.7025.6 / 29.7 / 33.80.262 / 0.287 / 0.309
162.12 / 2.50 / 1.6724.2 / 27.7 / 31.20.236 / 0.254 / 0.274

Q3: Can Gate3R handle priors estimated by real calibration, visual odometry, or depth estimation methods?

We evaluate on real priors from AnyCalib (intrinsics), DROID-SLAM (pose), and UniDepthV2 (depth), with heads and scores unchanged, under two deployment modes: (i) a reference-free fixed budget, and (ii) an Adaptive threshold set at the 95% score quantile of clean reference sets from the target domain, which adapts the discard rate to the data. Across all models, ρ = 0.3 beats both fixed policies on pose and depth by catching estimator failures, and Adaptive gains further on pose. AnyCalib intrinsics are already accurate on MapAnything, so keeping all priors is best there, with Adaptive close behind:

Prior sourceModelAllNoneρ = 0.3Adaptive
DROID-SLAM pose (AUC↑)MapAny 2v35.829.038.843.0
MapAny 4v34.229.436.339.6
Pow3R30.730.035.838.3
UniDepthV2 depth (inlier↑)MapAny 2v0.3940.3750.4030.396
MapAny 4v0.3370.3240.3410.332
Pow3R0.3960.4000.4090.406
AnyCalib K (°↓)MapAny 2v1.912.592.221.99
MapAny 4v1.872.562.181.94
Pow3R1.821.801.831.79

Q4: A fixed budget may discard accurate priors when corruption is rare or retain harmful priors when corruption is widespread; could an adaptive gating strategy help?

When corruption is rarer (clean fractions up to 1/2), ρ = 0.3 still beats both fixed policies on all axes but Pow3R depth. When it is widespread (1/6 clean, the rest at the two most severe levels), ρ = 0.3 discards too few, but a larger budget or an Adaptive threshold beats both fixed policies on most axes. The Adaptive threshold replaces ρ with the 95% quantile of fused scores on clean reference sets from the target domain, so the discard rate follows the data; without such references, ρ = 0.3 remains the default:

AllNoneρ = 0.3ρ = 0.5ρ = 0.7Adaptive
Pow3R K (°↓)3.951.823.042.512.101.98
Pow3R P (AUC↑)10.430.118.724.730.633.4
Pow3R D (inlier↑)0.4390.3940.4480.4430.4320.445
MapAny 2v K4.002.613.212.872.642.48
MapAny 2v P13.829.121.727.232.632.3
MapAny 2v D0.2870.3740.3620.3870.3900.364
MapAny 4v K4.002.583.162.762.582.24
MapAny 4v P12.429.420.725.929.929.5
MapAny 4v D0.2100.3230.2770.3120.3330.308
AuthorsOpenReview ↗

Official Response by Authors – Part 2

Q5: Attention entropy may reflect scene difficulty rather than prior corruption, and heads are selected by detection AUC rather than by their contribution to reconstruction; can these factors be disentangled?

The entropy level does track scene difficulty (error of the image-only reconstruction), but the shift caused by a corruption largely does not, and detection is unchanged among sets of equal difficulty. We select heads by detection AUC because it needs no ground-truth geometry; mild corruptions still rank low under the fused score, so Gate3R keeps mildly corrupted priors, and selecting heads on any other dataset gives almost the same results.

Dear Reviewer Gcpk,

As the discussion period ends, we’re following up on our rebuttal since we haven’t heard back. We value your insights and would appreciate it if you could review our responses and let us know whether they address your concerns.

Best regards, All authors

Thank you for the additional experiments and clarifications. The expanded multi-view evaluation and experiments with real estimated priors substantially address my main empirical concerns. I will raise my rating.

Reviewer_6KDTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Gate3R, a training-free method to improve the robustness of prior-conditioned feed-forward 3D reconstructions under noisy conditional inputs. The main idea is to check whether input priors, such as camera intrinsics, relative poses, and depth, are reliable. The authors propose to look at the difference between the input priors and model predictions, together with the models’ attention entropy to determine this. Based on these signals, Gate3R decides which priors to discard and reruns the reconstruction model without them. The paper also studies the echo effect, where the model tends to reproduce errors from incorrect priors in its predictions. Experiments on Pow3R and MapAnything across several datasets show that Gate3R generally performs better than always keeping or discarding the priors.

Critical strengths
  1. Simple and effective. The paper addresses an important problem in feed-forward 3D reconstruction, as inaccurate geometric priors can sometimes make the reconstruction worse. The proposed method is simple and does not require any additional training or fine-tuning, making it easy to apply to different reconstruction models.
  2. The paper presents an interesting observation about the echo effect, showing that reconstruction models tend to carry errors from input priors into their predictions. The authors also provide extensive experiments across multiple datasets, different corruption levels, and both single-prior and multi-prior settings.
Critical weaknesses
  1. Limited improvement over simpler baselines. While Gate3R generally performs better than confidence-based gating, the improvements are relatively small in several cases, especially for pose and depth estimation. Table 2 shows that confidence-based gating already achieves comparable results in some settings. Also, Table 5 shows that using only the prior residual gives very similar performance to the full method. This makes me question how much the attention entropy and score fusion actually help.
  2. Lack of evaluation with real-world geometric priors. Most experiments are based on synthetically corrupted geometric priors. I think it would be more convincing to directly use poses from off-the-shelf VO/SLAM methods or depth priors from monocular depth estimation models. These would better reflect the errors we might encounter in practice and help demonstrate how useful the method is in real-world applications.
  3. The method seems rather heuristic. I find several design choices somewhat heuristic, including the fixed 30% discard ratio, batch-level score normalization, and tail-weighted fusion of residual and attention entropy. It is unclear how well these choices generalize to different corruption rates and distributions, especially when most priors in a batch are severely corrupted.
Questions for authors

1. Comparison with confidence-based baselines. Table 5 shows that residual-only gating already achieves very similar performance to the full method. Could the authors explain in which cases attention entropy provides a clear benefit? It would also be helpful to see a comparison with residual-only gating in the multi-prior setting.

2. Evaluation with real-world priors. Could the authors evaluate Gate3R using geometric priors obtained from off-the-shelf VO/SLAM or monocular depth estimation methods? I think this would provide a more realistic evaluation than using synthetically corrupted ground-truth priors.

3. Robustness of the gating strategy. How sensitive is the method to the choice of the 30% discard ratio? In particular, how would it perform when most priors in a batch are severely corrupted, rather than having errors spread across different severity levels? Could the method adapt its discard ratio based on the actual reliability of the priors?

4. Echo effect. Were Pow3R and MapAnything trained with noisy geometric priors? I wonder whether the observed echo effect could be simply explained by a train-test distribution shift, especially if the models were mainly trained with clean priors.

4 comments
AuthorsOpenReview ↗

Official Response by Authors – Part 1

We sincerely thank the reviewer for the detailed and constructive comments. We take every comment seriously and hope our response addresses the reviewer's concerns. If any questions remain, we are happy to address them.

Q1: Residual-only gating performs close to Gate3R; when does attention entropy clearly help, and how do they compare in multi-prior settings?

(a) When entropy helps. Entropy helps most on Pow3R intrinsics, where it catches corruptions the residual misses. On Pow3R depth it costs slightly, since mildly corrupted depth priors still help.

(b) Multi-prior setting. At the same discard budget, Gate3R matches or beats residual-only gating in most cells, with the largest gain again on Pow3R K (column = evaluated prior / prior set; K ray err °↓, P AUC@5°↑, D inlier↑):

ModelMethodK/KPK/KDK/KPDP/KPP/PDP/KPDD/KDD/PDD/KPD
Pow3RResidual only1.6001.4821.52032.7632.8933.540.5520.5100.508
Gate3R1.5251.3341.39732.4832.9833.440.5620.5150.506
MapAny 2vResidual only1.7011.6741.71637.1537.0538.360.5160.4100.455
Gate3R1.6991.6771.69237.2837.5438.610.5170.4140.462
MapAny 4vResidual only1.6411.6651.74234.1935.2135.320.4320.3560.406
Gate3R1.6411.6411.71534.2535.4435.330.4360.3670.415

Q2: Could the authors evaluate Gate3R using geometric priors obtained from off-the-shelf VO/SLAM or monocular depth estimation methods?

We evaluate real priors from DROID-SLAM (pose) and UniDepthV2 (depth) with heads unchanged, under (i) a reference-free budget ρ = 0.3 and (ii) an Adaptive threshold at the 95% score quantile of clean target-domain reference sets (Q3c). On all models, ρ = 0.3 beats both fixed policies by catching failures; Adaptive gains further on pose:

Prior sourceModelAllNoneρ = 0.3Adaptive
DROID-SLAM pose (AUC↑)MapAny 2v35.829.038.843.0
MapAny 4v34.229.436.339.6
Pow3R30.730.035.838.3
UniDepthV2 depth (inlier↑)MapAny 2v0.3940.3750.4030.396
MapAny 4v0.3370.3240.3410.332
Pow3R0.3960.4000.4090.406

Q3: Sensitivity to the 30% discard ratio, performance under heavy batch corruption, adaptive discard rate, and heuristic design choices (batch normalization, tail-weighted fusion).

(a) Sensitivity to ρ. The paper sweeps ρ from 0 to 1 for clean fractions 1/6 to 1: with any corruption, Gate3R beats random discarding at every budget on pose and depth, and every ρ in [0.3, 0.5] beats both fixed policies on all 27 multi-prior cells. As the clean fraction is unknown at deployment, we fix ρ = 0.3.

(b) Heavy corruption. With heads fixed, 1/6 of the sets are clean and the rest carry the two most severe levels. ρ = 0.3 under-discards, mostly trailing None. ρ = 0.7 or the Adaptive threshold (c) beats both on most axes:

AllNoneρ = 0.3ρ = 0.5ρ = 0.7Adaptive
Pow3R K (°↓)3.951.823.042.512.101.98
Pow3R P (AUC↑)10.430.118.724.730.633.4
Pow3R D (inlier↑)0.4390.3940.4480.4430.4320.445
MapAny 2v K4.002.613.212.872.642.48
MapAny 2v P13.829.121.727.232.632.3
MapAny 2v D0.2870.3740.3620.3870.3900.364
MapAny 4v K4.002.583.162.762.582.24
MapAny 4v P12.429.420.725.929.929.5
MapAny 4v D0.2100.3230.2770.3120.3330.308

(c) Adaptive threshold. With clean target-domain reference sets (20 per dataset on the synthetic benchmark; held-out scenes for real priors), we replace ρ by the 95% quantile of their fused scores, so the discard rate follows the batch. Score scales shift across datasets, so thresholds do not transfer and ρ = 0.3 stays the reference-free default.

AuthorsOpenReview ↗

Official Response by Authors – Part 2

(d) Batch normalization and tail-weighted fusion. With heads and ρ = 0.3 fixed, we compute normalization statistics and rankings within random windows of B sets: accuracy barely changes down to B = 10. Replacing the tail weights by equal weights is never better and is worse on most axes, because a weak signal then dilutes the strong one:

AllNoneB = 10B = 20B = 50Full batch (Gate3R)Equal weights
Pow3R K (°↓)2.071.801.601.581.551.541.55
Pow3R P (AUC↑)24.330.031.832.132.232.132.0
Pow3R D (inlier↑)0.5900.4000.5550.5540.5560.5580.557
MapAny 2v K2.202.591.881.851.811.791.85
MapAny 2v P28.429.035.335.735.836.135.6
MapAny 2v D0.3930.3750.4300.4320.4380.4350.435
MapAny 4v K2.212.561.841.791.761.761.80
MapAny 4v P24.929.432.733.033.033.133.1
MapAny 4v D0.3010.3240.3460.3510.3540.3550.355

Q4: Were Pow3R and MapAnything trained with noisy priors, or could the echo effect simply be a train-test distribution shift?

Both are trained with clean priors, so noisy real priors are OOD. Yet OOD shift alone does not explain the echo pattern: MapAnything echoes pose more than K (92% vs. 79%), whereas Pow3R does the opposite (66% vs. 85%), suggesting echo depends on architecture and injection mechanism.

I would like to thank the authors for their detailed rebuttal and additional experiments. The real-world prior evaluation (Q2) and robustness analysis (Q3) largely address my concerns. The additional comparisons (Q1) are helpful, although the gains over residual-only gating remain modest in most settings. For Q4, the role of distribution shift in the echo effect remains somewhat unclear. I appreciate the authors' efforts and encourage them to include these results and discussions in the revised paper.

Thank you for the thoughtful feedback and for recognizing our real-prior and robustness results!

Regarding the echo effect, we agree that the role of distribution shift remains open. Both models are trained with clean priors, and isolating it would require retraining with noisy priors, which we leave to future work. Gate3R targets released models as they are and needs neither retraining nor the cause of the echo. We will clarify this in the final version.

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

This paper presents Gate3R, a training-free method for deciding whether to use geometric priors (camera intrinsics, pose, and depth) in prior-conditioned feed-forward 3D reconstruction. The authors highlight an “echo effect,” in which the reconstruction model reproduces errors present in the injected priors. Gate3R avoids this effect by keeping the prior-conditioned reconstruction only when the priors are judged reliable; otherwise, it performs a second feed-forward pass using the images alone. To assess reliability, the method fuses two signals obtained from the prior-conditioned pass: (1) the residual between the injected priors and the model’s predictions, and (2) the attention entropy of a selected transformer head. The method is evaluated across six datasets, three prior types, and two reconstruction models (Pow3R and MapAnything), in both single- and multi-prior settings, and shows consistent improvements over both always using and always discarding the priors.

Critical strengths
  • The problem tackled by this paper is sound, and the proposed approach is elegant and effective. There is no need for retraining or fine-tuning the feed-forward reconstruction models.

  • The identification and analysis of the "echo effect" is an interesting contribution of the paper.

Critical weaknesses
  • The method relies on a fixed discard budget ρ=0.3\rho = 0.3. The optimal budget varies significantly with corruption frequency, as shown in the paper. In real setting the corruption rate is unknown, making it unclear why the discard budget should be fixed.

  • The paper is not always clear, some sections need a lot of effort to be understood.

  • The baselines are limited to 4 basic selection policies: confidence-based selection, random selection, always use priors, and always discard priors. More sophisticated approaches could be also added to the comparison to make the paper more strong. As an example, the authors could train a network for prior reliability estimation.

  • The experiments rely on synthetic corruptions of the priors; adding tests with realistic corruptions, such as noise from actual sensors or estimation pipelines, would strengthen the evaluation.

Minor comments:

  • In lines 149 and 156, the same letter dd is used for two different notions.
  • The Figure 3 could be improved by using a visualization of the residual between the corrupted and the clean attention entropy maps.
Questions for authors
  • How could one choose the discard budget ρ\rho in realistic deployments?

  • Can Gate3R be used with weighting priors according to reliability rather than using a binary keep/discard decisions? btw, this solution can handle the heuritic way to choose the discard budget!

3 comments
AuthorsOpenReview ↗

Official Response by Authors

We sincerely thank the reviewer for the detailed and constructive comments. We take every comment seriously and hope our response addresses the reviewer's concerns. If any questions remain, we are happy to address them.

Q1: How could one choose the discard budget ρ in realistic deployments? Adding tests with realistic corruptions would strengthen the evaluation.

(a) Choosing ρ. Because the corruption rate is unknown and score scales vary across domains, Gate3R supports two deployment modes: (i) a reference-free fixed budget (ρ = 0.3), and (ii) an Adaptive threshold set at the 95% score quantile of clean reference sets from the target domain, which adapts the discard rate to the data.

(b) Realistic corruptions. Keeping the same settings, we evaluate both modes on real priors from DROID-SLAM (pose) and UniDepthV2 (depth). Across all models, ρ = 0.3 beats both fixed policies on pose and depth by catching estimator failures, while Adaptive gains further on pose via a data-driven discard rate:

Prior sourceModelAllNoneρ = 0.3Adaptive
DROID-SLAM pose (AUC↑)MapAny 2v35.829.038.843.0
MapAny 4v34.229.436.339.6
Pow3R30.730.035.838.3
UniDepthV2 depth (inlier↑)MapAny 2v0.3940.3750.4030.396
MapAny 4v0.3370.3240.3410.332
Pow3R0.3960.4000.4090.406

Q2: Can Gate3R be used with weighting priors according to reliability rather than using a binary keep/discard decisions?

We scale each prior's embedding by a weight w∈[0,1]w \in [0,1] without fine-tuning: w=1w = 1 keeps the prior and w=0w = 0 exactly removes it. Partial weights shift poses and Pow3R depth off the training distribution: with w=0.5w = 0.5, pose AUC drops to 10–13, below both keeping and discarding the prior. Training-free weighting by the Gate3R score ss (w=1−s/3w = 1 - s/3) therefore helps intrinsics only, so we keep the binary decision:

MethodMapAny 2v K↓MapAny 2v P↑MapAny 2v D↑MapAny 4v K↓MapAny 4v P↑MapAny 4v D↑Pow3R K↓Pow3R P↑Pow3R D↑
All2.2028.430.3932.2124.910.3012.0724.260.590
None2.5929.040.3752.5629.430.3241.8030.020.400
w=0.5w = 0.52.3012.830.3672.2610.410.3101.8812.320.248
Score-weighted1.6333.450.4341.6030.740.3521.4830.570.538
Gate3R1.7936.050.4351.7633.130.3551.5432.120.558

Q3: More sophisticated approaches could be also added to the comparison, the authors could train a network for prior reliability estimation.

We train a linear probe on the same per-set signals to predict whether the prior is corrupted, using LaMAR with the same synthetic corruptions as the benchmark, and use its probability in place of the fused score (same budget ρ = 0.3). Gate3R matches or beats this supervised probe without any labels, while the probe needs labels and retraining for each model and prior. Despite being trained in distribution, the probe trails Gate3R on 7 of 9 axes:

MethodMapAny 2v K↓MapAny 2v P↑MapAny 2v D↑MapAny 4v K↓MapAny 4v P↑MapAny 4v D↑Pow3R K↓Pow3R P↑Pow3R D↑
All2.2028.430.3932.2124.910.3012.0724.260.590
None2.5929.040.3752.5629.430.3241.8030.020.400
Linear probe1.8435.950.4291.7833.060.3491.5832.250.563
Gate3R1.7936.050.4351.7633.130.3551.5432.120.558

Q4: Some sections lack clarity; minor issues with duplicate notation dd (L149, L156) and Fig. 3 entropy visualization.

We will revise the method section for clarity, and have fixed both minor comments in the revision.

Reviewer_NLYVOpenReview ↗

Official comment of Reviewer NLYV

I thank the authors for their detailed response, I will maintain my score. I suggest that the authors use the remaining space in the paper (page 10) to include more results from the appendix, particularly those that address the reviewers’ concerns.

Thank you for your feedback and the helpful suggestion! We will reorganize the paper in the revision.

Reviewer_o7M4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Gate3R decides which geometric priors to discard in frozen feed-forward reconstruction models. It combines the prior–prediction residual with attention entropy, ranks image sets within a deployment batch, and reruns a fixed fraction without the flagged priors. Experiments on Pow3R and MapAnything show improvements under controlled corruptions, including cases with several priors. The paper also measures how strongly the models copy errors from their inputs.

Critical strengths
  1. The problem is useful and clearly isolated. Geometric conditioning can make a model confidently reproduce a bad prior. The echo measurements in Table 4 explain why output confidence alone is insufficient, and the proposed gate can be added to existing checkpoints.
  2. The controlled evaluation is thorough. Tables 2–3 compare paired policies across models and prior combinations, while the appendix checks head-selection transfer, corruption rates, and discard budgets. I also appreciate the explicit Pow3R depth failure and the per-dataset breakdown rather than only reporting aggregate gains.
Critical weaknesses
  1. Deployment robustness remains uncertain. All corrupted priors come from prescribed perturbations of ground truth. The fusion and ranking also use every view set of a dataset as one batch (§4.1). The corruption-rate sweeps are helpful, but do not establish stability with small batches or the correlated, spatially varying errors of actual pose/depth estimators. These settings matter because the selected priors depend on the batch's score distribution.
  2. The added value of entropy is modest in most settings. In Table 5, residual-only gating already obtains nearly all the MapAnything gains: for two views, ray error is 1.80 versus 1.79 and pose AUC is 35.9 versus 36.1. Some axes favor the residual alone. The max-clean-z fusion control is useful, but only means over ten corruption seeds are reported. Paired uncertainty would help establish how much entropy selection and tail weighting add beyond the strong residual-gating baseline.
Questions for authors
  1. Can you report results using priors from an actual pose or depth estimator, and quantify sensitivity to deployment batch size while keeping heads and the discard budget fixed?
  2. Can you report paired uncertainty across the ten corruption seeds for the differences from residual-only gating in Table 5? Would equal-weight fusion of standardized residual and entropy give comparable results?
3 comments
AuthorsOpenReview ↗

Official Response by Authors

We sincerely thank the reviewer for the detailed and constructive comments. We take every comment seriously and hope our response addresses the reviewer's concerns. If any questions remain, we are happy to address them.

Q1: Can you report results using priors from an actual pose or depth estimator, and quantify sensitivity to deployment batch size while keeping heads and the discard budget fixed?

(a) Priors from real estimators. We evaluate on real priors from DROID-SLAM (pose) and UniDepthV2 (depth) under two deployment modes: (i) a reference-free fixed budget (ρ = 0.3), and (ii) an Adaptive threshold set at the 95% score quantile of clean reference sets from the target domain. Across all models, ρ = 0.3 beats both fixed policies on pose and depth by catching estimator failures, while Adaptive gains further on pose:

Prior sourceModelAllNoneρ = 0.3Adaptive
DROID-SLAM pose (AUC↑)MapAny 2v35.829.038.843.0
MapAny 4v34.229.436.339.6
Pow3R30.730.035.838.3
UniDepthV2 depth (inlier↑)MapAny 2v0.3940.3750.4030.396
MapAny 4v0.3370.3240.3410.332
Pow3R0.3960.4000.4090.406

(b) Small deployment batches. With heads and ρ = 0.3 fixed, we randomly shuffle the test sets and compute normalization statistics and rankings within consecutive windows of size B. Accuracy barely changes down to B = 10: strongly corrupted sets remain in the top ρ, while small windows only perturb borderline sets near the cutoff:

AllNoneB = 10B = 20B = 50Full
Pow3R K (°↓)2.071.801.601.581.551.54
Pow3R P (AUC↑)24.330.031.832.132.232.1
Pow3R D (inlier↑)0.5900.4000.5550.5540.5560.558
MapAny 2v K2.202.591.881.851.811.79
MapAny 2v P28.429.035.335.735.836.1
MapAny 2v D0.3930.3750.4300.4320.4380.435
MapAny 4v K2.212.561.841.791.761.76
MapAny 4v P24.929.432.733.033.033.1
MapAny 4v D0.3010.3240.3460.3510.3540.355

Q2: Can you report paired uncertainty across the ten corruption seeds for the differences from residual-only gating in Table 5? Would equal-weight fusion of standardized residual and entropy give comparable results?

Paired uncertainty and equal-weight fusion. Below we report the difference to residual-only gating with paired 95% t-CIs over the 10 seeds (Δ > 0: better; bold: significantly better), together with equal-weight fusion (wres=went=1w_\text{res} = w_\text{ent} = 1):

ModelPriorResidual onlyGate3RΔ [95% confidence interval]Equal weights
Pow3RK (ray err °↓)1.631.54+0.094 [+0.072, +0.116]1.55
MapAny 2vK1.801.79+0.012 [+0.004, +0.020]1.85
MapAny 4vK1.771.76+0.013 [+0.006, +0.021]1.80
Pow3RP (AUC@5°↑)32.332.1−0.16 [−0.39, +0.07]32.0
MapAny 2vP35.936.1+0.15 [+0.01, +0.29]35.6
MapAny 4vP33.133.1+0.03 [−0.05, +0.11]33.1
Pow3RD (inlier↑)0.5610.558−0.003 [−0.005, −0.001]0.557
MapAny 2vD0.4330.435+0.002 [0.000, +0.004]0.435
MapAny 4vD0.3500.355+0.005 [+0.002, +0.007]0.355

When entropy helps. Gate3R beats residual-only gating on most axes. Entropy helps most on Pow3R intrinsics, where it catches corruptions the residual misses. On Pow3R depth it costs slightly, since mildly corrupted depth priors still help.

Equal-weight fusion. Equal weights at best match Gate3R, because giving a weak readout the same weight as the strong one adds noise to the fused score.

Reviewer_o7M4OpenReview ↗

Official Response to Authors

Thanks for the response. My main concerns have been addressed, so I’ll keep my weak accept score.

Thank you for the follow-up! We are glad our response resolved your concerns, and we appreciate your constructive feedback in improving our work.

Comments on the paper

2 comments

Dear Reviewers,

             The authors posted their comments in response to the reviews. Please make an effort to discuss your review and the authors' response. - regards AC
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive reviews! Reviewers found the problem sound and practical (NLYV, o7M4, 6KDT, Gcpk), the training-free approach elegant and effective (NLYV, 6KDT, Gcpk), the echo effect interesting (NLYV, 6KDT), and the evaluation thorough (o7M4, Gcpk).

We have updated the manuscript with the following revisions based on the reviewers' suggestions. Some results appear only in our responses to the reviewers. We will include them in the final version. All revisions in the updated version are highlighted in blue:

  1. Add more experiments:

    • Add evaluation with real estimator priors under both the fixed budget and an adaptive threshold (o7M4-Q1, NLYV-Q1, 6KDT-Q2,Q3, Gcpk-Q3,Q4).
    • Add multi-prior residual-only results and equal-weight fusion (o7M4-Q2, 6KDT-Q1,Q3).
    • Add robustness tests: small deployment batches, heavy corruption and sensitivity (o7M4-Q1, 6KDT-Q3, Gcpk-Q4).
    • Add a supervised linear probe as a baseline (NLYV-Q3).
    • Add results with 8 and 16 input views (Gcpk-Q2).
    • Add layer-wise entropy analysis (Gcpk-Q1).
  2. Add more discussion and details:

    • Discuss how to set the discard rule in deployment (NLYV-Q1, 6KDT-Q3, Gcpk-Q4).
    • Discuss when attention entropy helps beyond the residual (o7M4-Q2, 6KDT-Q1).
    • Fix duplicate notation and add an entropy difference map to Fig. 3 (NLYV-Q4).
    • Show that the entropy reads prior–image disagreement (Gcpk-Q1).

Thanks again for all the effort and time, and we look forward to further discussion if there are any more questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.