acceptodds
Under review as a conference paper at ICLR 2027

When Absence Is Not Evidence: Occlusion-Type-Aware Censored Fusion for Multi-View 3D Grounding

Abstract

In multi-view 3D grounding, a low language-conditioned response can indicate that a surface does not belong to the referent or that it is occluded. Conflating these cases can dilute valid evidence or accumulate spurious contradiction. We formulate fusion as censored observation modeling, marginalizing uncertain track–view validity. Exact visibility labels reveal a readout-dependent distinction: partial occlusion can retain semantic evidence through spatially extended response fields, whereas full occlusion carries much weaker membership signal. Our three-component likelihood preserves partial-occlusion evidence and approximates full occlusion with a shared nuisance density. A spatial gate removes instance-grouping labels from inference while geometry and semantic backbones remain frozen. We introduce VisCen3D, a diagnostic benchmark with persistent surface identity, exact visibility, and occlusion-type labels. A preregistered evaluation on 160 fresh scenes and 882 episodes shows that grouping-free fusion on analytic tracks improves AUROC on the most-occluded surface tercile by under Stable Diffusion and under SigLIP over development-selected non-censored comparators; scene-level confidence intervals exclude zero. A post-hoc ablation on the same partition shows that the partial/full refinement improves over ordinary censoring. Top-1 target-track selection increases by / percentage points under SD/SigLIP, with the SD directional prediction confirmatory and the SigLIP result descriptive. Transfer experiments expose the limits: averaging outperforms every tested likelihood rule on ScanNet++ with windowed responses, while incomplete surface recovery limits fusion with fully estimated geometry. These findings establish ranking and selection benefits on controlled surface support and clarify dependence on semantic readout and surface recovery.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-view 3D visual grounding, visibility and occlusion reasoning, censored observation modeling, vision-language models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_DAJwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission studies how to fuse language-conditioned responses across views when a projected 3D surface may be occluded. The main observation is that partial occlusion can retain useful semantic evidence through spatially extended response maps, whereas fully occluded responses carry much weaker membership information. The proposed fusion rule combine class-conditional visible, and partially occluded response densities with a shared full-occlusion density, using estimated visibility and a local spatial gate that avoid instance-grouping labels at inference. The work also introduces VisCen3D for surface-level visibility diagnosis. On fresh synthetic scenes with analytic surface tracks, the method improves occluded-surface ranking and target-track selection over development-selected baselines. The real-data and estimated-geometry experiments are less favorable because the gains depend strongly on the response readout, and incomplete surface recovery substantially limits the practical benefit.

Critical strengths
  1. The partial/full distinction is motivated and supported by relevant results.

The submission goes beyond treating every occluded response as missing information. The measured difference between partial- and full-occlusion discrimination provide a concrete reason to retain the former. The resulting likelihood is simple enough that its behavior at the limiting cases can be understood directly.

  1. The comparisons isolate the fusion question reasonably well.

Tracks and cached semantic responses are held fixed, and the experiments include ordinary censoring, validity-weighted likelihood ratios, and hard censoring. The additional two-component comparison on Confirmation 2 is particularly useful, which shows that ordinray censoring alone does not explain the reported improvement.

  1. The benchmark supports a useful diagnostic task.

Persistent surface identity, geometric visibility checks, and partial/full labels allow questions that object-level grounding annotations cannot directly answer. The authors also take care not to present target-track selection as full object localization, or recovered-support ranking as complete surface coverage.

Critical weaknesses
  1. The strongest evidence remains confined to synthetic scenes with analytic support.

The primary gains assume that the relevant surfaces and their projections are already available. In the fully estimated setting, target-track recall is only 0.39, and the coverage-aware evaluation report differences of roughly 0.003 between fusion rules. On ScanNet++, windowed averaging is better than every tested likelihood rule, which do not invalidate the controlled finding, but they leave the practical grounding contribution substantially less established than the diagnostic contribution.

  1. The comparison with learned fusion weakens the case for the proposed model beyond a low-data setting.

At full supervision, the additive MLP is ahead in 5/10 SigLIP runs and 7/10 SD runs, with most remaining runs unresolved. Therefore, the structured model’s demonstrated advantage is mainly in the very-low-data regime, where it also fails to fit on two of five one-scene subsets. This is acknowledged, but the main experimental presentation should give these results more prominence, and reports absolute performance alongside the likelihood baselines.

  1. The spatial gate is useful empirically but insufficiently characterized.

Taking the maximum estimated visibility within a radius tied to the entire track-set extent can admit evidence from neighboring objects, and may be sensitive to scene scale, track density, and outliers. Non-inferiority to the instance-grouped gate on the synthetic confirmation set do not establish robustness to these factors. A radius sweep and comparisons with simpler gates would help to establish whether the spatial construction itself is necessary.

  1. The partial/full ablation does not fully isolate the source of the gain.

The refinement changes the response densities, which introduces class-conditional censored evidence, and add a spatial gate. Its improvement over the two-component model establishes the benefit of the complete refinement, but not whether accurate partial/full separation is essential. Constant-gate and shuffled-gate controls would help to distinguish meaningful spatial conditioning from the benefit of a more flexible response model.

Questions for authors
  1. What exactly supplies the validity cues in the analytic-track experiments?

Please specify which quantities come from analytic geometry, and which come from estimated geometry, including the construction of depth margins and point-map residuals. I suggest that a dependency table would make the test-time assumptions easier to assess.

  1. How much does the spatial gate contribute beyond a globally fitted mixture?

Could the authors compare the current gate with a development-selected constant ρ\rho, ρ=η\rho=\eta, and a shuffled spatial gate, while keep the response densities fixed? How sensitive are the results to the neighborhood radius?

  1. Is the shared full-occlusion density preferable to a regularized class-conditional alternative?

Full-occlusion AUROC of 0.45 under SigLIP suggests some inverse association, even if weak. A comparison with separately fitted full-occlusion densities would clarify whether sharing hh is beneficial regularization, or it may discard useful information.

4 comments
AuthorsOpenReview ↗

Gate controls, learned fusion comparisons, and geometric cue dependencies

Thank you for the suggestions. New Confirmation 2 analyses use the same 160 scenes/882 episodes per backbone and cached inputs. They are post hoc; intervals are paired scene-bootstrap 95%. All AUROCs are AUROC_occ; paired entries are SD / SigLIP. Differences use unrounded metrics.


W1. Scope. The strongest evidence concerns controlled surface ranking and target-track selection. Estimated geometry has target-track recall 0.39 and coverage-aware fusion differences near 0.003. We will foreground these limits and the readout dependence of real-data results.


W2. Learned fusion. App. D uses the 120-scene/672-episode core population and oracle-grouped ρ. The new spatial comparison uses the submitted gate during both training and evaluation on the same population.

Dev scenes / gateSD: mixture / MLPSigLIP: mixture / MLP
1 / grouped, 3/5 subsets fittable0.864 / 0.8010.913 / 0.881
5 / grouped0.871 / 0.8550.927 / 0.928
48 / grouped0.871 / 0.876±0.0050.932 / 0.934±0.001
48 / spatial0.872 / 0.877±0.0030.926 / 0.929±0.001

Subset rows are means; full-data MLP entries are mean±SD over 10 seeds. The best non-censored core rules reach 0.824 / 0.883. The MLP has higher full-data means. The low-data evidence uses grouped ρ and is conditional on fitting success: two of five one-scene subsets fail. The new full-data run does not establish a spatial low-data advantage. We will foreground these absolute results.


W3–W4/Q2. Gate controls and radius. Three-component controls share f0,f1,g0,g1,hf_0,f_1,g_0,g_1,h and all other inputs; only ρ changes. The two-component reference retains its pooled censored density c.

Gate/modelAUROC_occR@1 (%)
Two-component0.817 / 0.86664.97 / 71.77
ρ=0, retaining hh0.844 / 0.87565.87 / 71.66
ρ=10.828 / 0.89855.44 / 60.77
Dev-selected constant0.859 / 0.91764.74 / 67.12
Shuffled spatial gate0.847 / 0.90367.03 / 71.72
Spatial, γ=0.06\gamma=0.060.868 / 0.92767.23 / 74.04

The constant gate recovers most of the AUROC gain over two-component censoring, but its R@1 point estimates are lower. Spatial-minus-constant AUROC is about +0.010 [+0.007, +0.014] for each backbone; R@1 gains are +2.5 pp [+0.7, +4.4] / +6.9 pp [+4.7, +9.2].

The shuffle uniformly permutes whole per-track gate vectors within each scene, without stratification (five permutations; metrics averaged), holding η, frustum masks and densities fixed. It preserves each view's raw gate distribution over all tracks, but changes its alignment with validity and scored observations; ρ≥η is not reimposed. The 0.022 / 0.024 AUROC loss therefore reflects this combined disruption, rather than isolating spatial conditioning.

AUROC varies by about 0.003 across γ=0.06–0.10\gamma=0.06–0.10 and decreases at smaller or much larger radii. At γ=0.06\gamma=0.06, neighborhoods average 39 tracks, 6% from other instances. These measurements do not establish robustness to changed track density, outliers or scene composition.


Q1. Cue dependencies. Ground truth supplies analytic tracks x_t, cameras and projections. VGGT supplies depth D̂_i, depth confidence, point maps P̂_i and predicted camera extrinsics. Let S be the similarity fitted from VGGT camera centers to ground-truth centers. We use its inverse, X̃_t=S⁻¹(x_t), to express each analytic track in VGGT's frame; point maps are unchanged.

Let p_ti be the nearest-pixel ground-truth projection, and z_ti the depth of X̃_t under VGGT's predicted extrinsics for view i. P̂_i(p_ti) and X̃_t use the same reconstruction frame. For positive z_ti, the raw cues are m_ti=(D̂_i(p_ti)−z_ti)/z_ti and r_ti=‖P̂_i(p_ti)−X̃_t‖₂/z_ti. A negative margin means that the sampled predicted depth is smaller than the track's estimated camera depth.

A logistic predictor uses these cues and depth confidence, with clipping and development-set median/IQR standardization; a histogram calibrator yields η. Both models are fitted on labeled development scenes. Primary spatial inference uses no test validity, membership, type or instance-group labels; oracle-grouped diagnostics use instance groups. Estimated validity still depends on analytic geometry. Fully estimated inference uses VGGT tracks and cues in its own frame; ground truth supplies development supervision and evaluation labels, not test-time inputs.


Q3. Shared full-occlusion density. With bins and other components fixed, we fitted separate h0,h1h_0,h_1 and interpolated toward shared hh. Shared-minus-separate AUROC gains are +0.0025 [+0.0015,+0.0037] / +0.023 [+0.014,+0.037]; under SigLIP, full-occlusion negative evidence rises from 0.016 to 0.199 nats with separate fits. The shared endpoint performs best among the tested settings, supporting regularization within this density family rather than absence of usable full-occlusion information.

Thank the author for the careful responses and more experiments. I have carefully read your rebuttal. Most of my key concerns are addressed, like the geometric cue dependencies, the rationale for sharing the full-occlusion density, and the contribution of the spatial gate beyond a constant mixture.

While some concerns remain regarding the limited evidence under estimated geometry, and the scope of the low-data advantage. In particular, the reported low-data results use the instance-grouped gate and are conditional on successful fitting. Is it possible to provide a corresponding low-data comparison using the grouping-free spatial gate, reporting both performance and fitting success rates? which help to clarify whether the data-efficiency benefit extends to the proposed inference setting.

Despite these remaining limitations, I think that the occlusion analysis and the diagnostic benchmark are valuable contributions, and the rebuttal has addressed most of my concerns. Thus, I raise my rating from 2 to 3.

AuthorsOpenReview ↗

Additional Low-Data Results with the Grouping-Free Spatial Gate and Fitting Success Rates

Reply to the Area Chair and Reviewer DAJw: spatial-gate low-data comparison and fitting success

Thank you for identifying the remaining question. Summary: with the grouping-free gate, the one-scene mixture advantage is +0.032 (SD) and +0.012 (SigLIP), with 46/48 mixture fits; SigLIP results vary substantially across development scenes. We ran a post-hoc low-data comparison with the submitted grouping-free spatial gate (γ=0.06) on the Appendix D population (48 development scenes; 120 test scenes/672 episodes; analytic tracks). Both models use the spatial gate in evaluation, and the MLP uses it during training. The original oracle-grouped results were reproduced on their original subsets before this extension.

We evaluated all 48 possible one-scene development subsets, 30 sampled subsets each at 2 and 5 scenes, and the original five subsets each at 10 and 25 scenes. The MLP uses 3 training seeds per subset at 1–5 scenes and 5 seeds at 10–25 scenes; the 48-scene row uses the previously reported 10-seed comparison. Entries are mean AUROC_occ over subsets where both models can be evaluated, with MLP scores averaged over seeds. Differences use unrounded scores. The final column counts subsets whose mixture-minus-MLP paired 95% test-scene-bootstrap intervals are positive / overlap zero / negative. These exploratory counts are not multiplicity-adjusted.

Stable Diffusion (SD)

Dev scenesMixture fittedMixture / MLPDifferenceMixture ahead / unresolved / MLP ahead
146/480.856 / 0.825+0.03240 / 6 / 0
230/300.865 / 0.837+0.02827 / 3 / 0
530/300.868 / 0.853+0.01526 / 4 / 0
105/50.871 / 0.863+0.0083 / 2 / 0
255/50.872 / 0.868+0.0042 / 3 / 0
481/10.872 / 0.877±0.003−0.005—

SigLIP

Dev scenesMixture fittedMixture / MLPDifferenceMixture ahead / unresolved / MLP ahead
146/480.899 / 0.888+0.01227 / 10 / 9
230/300.911 / 0.901+0.01016 / 11 / 3
530/300.919 / 0.921−0.0025 / 14 / 11
105/50.924 / 0.925−0.0010 / 4 / 1
255/50.924 / 0.927−0.0030 / 1 / 4
481/10.926 / 0.929±0.001−0.003—

Fitting success and its denominator. The mixture fits on 46/48 (95.8%) single-scene choices; the MLP trains on 48/48. The two mixture failures are the two scenes containing no fully occluded observations, leaving the shared full-occlusion density h without fitting data. Both failures happened to occur in the paper's original five one-scene draws, yielding its unrepresentative 3/5 fit count. For those two scenes, the MLP scores 0.809/0.812 (SD) and 0.865/0.870 (SigLIP); these are not included in the paired means. All evaluated subsets at 2, 5, 10 and 25 scenes fit successfully. A separate fit-only screen of 200 sampled subsets at each of 2, 5 and 10 scenes also found 200/200 valid mixture fits per size; these extra subsets are not included in the performance tables. The reported one-scene performance means are conditional on the 46 jointly evaluable subsets, not on all 48.

Interpretation. The low-data advantage survives removal of oracle instance grouping, but is backbone-dependent. Under SD, the mixture leads on 40/46 one-scene and 26/30 five-scene subsets, and is never significantly behind on any evaluated subset with 1–25 scenes. Its mean advantage decreases with more supervision; the MLP is slightly ahead at 48 scenes. Under SigLIP, the mean one-scene margin is only +0.012, with the mixture ahead on 27/46 subsets and the MLP ahead on 9/46; the MLP's mean overtakes by five scenes. On the paper's original three fittable one-scene subsets, SigLIP's margin was +0.032 with the grouped gate (MLP seed 0) and +0.031 with the spatial gate (also +0.031 using seed 0 alone or averaging seeds 0–2, as in the table). In the three-seed spatial analysis, these subsets ranked 6th, 9th and 10th among 46 by margin. Their larger gains than the 46-subset mean reflect which random scenes were drawn, not evidence of a substantial grouped-gate advantage.

We will foreground these spatial-gate comparisons and fitting denominators, retain the oracle-grouped results as diagnostics, replace the paper's "ahead in every repeat" claim with the subset counts, qualify the single-scene fit requirement, and acknowledge that the MLP is competitive or slightly better at full supervision. These are exploratory analyses on synthetic analytic support and a previously examined test split; per-subset bootstrap intervals quantify test-scene sampling uncertainty, not uncertainty over development-subset choice or training seeds.

Dear authors, thank you for your detailed responses. The low-data comparisons using the grouping-free spatial gate, together with the fitting success rates, answer my remaining question. The reporting of the backbone-dependent gains and fitting limitations is good. I hope these qualifications and the complete results can be clearly reported into the revised manuscript.

Reviewer_H7dnOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies multi-view 3D grounding under uncertain surface visibility. The authors formulate multi-view semantic fusion as censored observation modeling, marginalizing uncertainty about whether a track-view response observes the intended surface. Based on the observation that partial occlusion can retain semantic evidence while full occlusion carries much weaker membership signal, the method uses a three-component likelihood to distinguish visible, partially occluded, and fully occluded responses. Class-conditional densities are used for visible and partially occluded responses, while full occlusion is modeled with a shared nuisance density. A grouping-free spatial gate is introduced to estimate the partial/full mixture without instance-grouping labels at inference. The paper also introduces VisCen3D, which provides persistent surface identity, exact per-view validity, and occlusion-type labels. Experiments on synthetic and real datasets evaluate ranking, target-track selection, transfer across semantic readouts, and fully estimated geometry.

Critical strengths
  1. The paper identifies a meaningful ambiguity in multi-view semantic fusion under occlusion and provides a principled formulation for addressing it. In particular, the measured distinction between partial and full occlusion directly motivates the three-component censored likelihood, rather than treating all occluded responses as equally uninformative. This problem formulation and the corresponding observation model constitute the main conceptual contribution of the paper.
  2. The experimental design provides strong support for the proposed analysis. The main evaluation uses fresh scenes, preregistered analysis rules, frozen comparators, and paired confidence intervals, while additional experiments examine transfer to Replica and ScanNet++, as well as fully estimated geometry. The inclusion of both successful and unsuccessful transfer cases makes the empirical conclusions more credible and helps delineate the conditions under which the proposed fusion model is effective.
Critical weaknesses

The benefit of the proposed censoring model appears to depend on the semantic readout and fitted response model. On Replica, the ranking gain over the strongest non-censored rule remains unresolved, while on ScanNet++ with windowed responses, mean-all outperforms every tested likelihood rule. The relative ordering reverses with mask-crop responses. Although the paper identifies a readout–response-model interaction, the current analysis does not isolate which factors are responsible for this reversal, making it difficult to determine the conditions under which the proposed type-aware fusion provides an advantage over simpler fusion rules.

Questions for authors

Please refer to the concern raised in the Critical Weaknesses section and provide further clarification on the observed readout–response-model interaction.

1 comment
AuthorsOpenReview ↗

Readout dependence, evidence transformation, and type-aware fusion

Thank you for highlighting the readout–response-model interaction. We ran post-hoc diagnostics on the ScanNet++ partition of App. F.3 (220 region-scenes, 900 episodes), first reproducing its results. Our model here uses ground-truth instance groups with estimated validity; this is an oracle-grouped diagnostic. All ranking values below are AUROC_occ; intervals are paired scene-bootstrap 95% intervals.


1. The evidence transformation affects the relative ordering.

Fusion ruleWindowedMask-crop
Mean-all0.7620.544
Sum of fixed response logits0.7850.610
Sum of fitted LLRs, constrained monotone0.7070.622
Sum of fitted LLRs0.7010.664
Type-aware fusion0.7260.696

The first four rules ignore validity; our model is included for comparison. The fixed-logit and fitted-LLR rules sum over the same in-frustum observations. Fitted LLR minus fixed logit is −0.084 [−0.097, −0.072] for windows but +0.054 [+0.037, +0.071] for crops. The readouts therefore favor different transformations before summation. The logit is a fixed comparison transform; it does not assume calibrated membership probabilities.

Constraining the fitted LLR to be monotone reduces crop AUROC from 0.664 to 0.622 (−0.042 [−0.051, −0.034]). This supports the usefulness of the unconstrained transformation in this experiment. It does not establish the population response–membership relationship or explain why the windowed fitted mapping underperforms.

Windowed fixed-logit fusion exceeds our model by +0.058 [+0.045, +0.072]. Switching from windows to crops also lowers our absolute AUROC (0.726 to 0.696). Thus the crop reversal concerns the relative ordering of fusion methods.


2. Tested estimation interventions do not fully close the gap. Jointly replacing estimated validity and type with exact evaluation labels changes our model by +0.003 [−0.006, +0.012] for windows and +0.005 [−0.003, +0.013] for crops. This joint intervention does not isolate validity estimation alone.

A separate diagnostic fits densities on other test scenes and evaluates held-out scenes, changing AUROC by +0.007 [+0.002, +0.013] and +0.009 [+0.003, +0.014]. This uses test-partition labels for diagnostic fitting; it is separate from the original development-only procedure. Reducing histogram resolution from 48 to 12 bins changes AUROC by −0.002 [−0.006, +0.002] / −0.007 [−0.012, −0.002]. Using exact validity and cross-fitted densities together still trails windowed mean-all by −0.016 [−0.028, −0.004]. These tested changes are insufficient to close the gap; other density errors or interactions remain possible.


3. What remains supported for type-aware censoring. Against fitted non-censored LLR, type-aware fusion improves by +0.026 [+0.016, +0.036] for windows and +0.032 [+0.023, +0.041] for crops. Against ordinary censoring, the differences are +0.009 [+0.005, +0.014] and +0.004 [+0.000, +0.008]; the crop gain is small and its displayed lower endpoint rounds to zero. These results support the complete refinement within the evaluated likelihood family, while the fixed-logit baseline remains stronger for windows.

Replica's original ranking contrasts remain unresolved: −0.024 [−0.056, +0.006] for SigLIP and −0.034 [−0.070, +0.002] for SD. The ScanNet++ diagnostics do not establish the cause of Replica's outcome.


4. A development-side diagnostic. Cross-fitting on the 57 development scenes (244 episodes) gives fitted-LLR-minus-fixed-logit differences of −0.089 [−0.111, −0.066] for windows and +0.069 [+0.041, +0.096] for crops, agreeing in sign with the test results. This comparison uses development labels, but was examined post hoc. It is a useful check in these two readouts, not a validated selection rule or a necessary/sufficient condition for success on new domains.


We will add the fixed-logit baseline, absolute crossed-readout results and gate assumptions, and distinguish measured sensitivity to the evidence transformation from an established causal explanation.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Attribution of gains: matched spatial baselines, MLP ablations, and evaluation scope

Thank you for identifying the attribution gap. New Confirmation 2 analyses use the same 160 scenes/882 episodes per backbone and cached inputs. They are post hoc; intervals are paired scene-bootstrap 95%. All AUROCs are AUROC_occ; paired entries are SD / SigLIP, and R@1 is in percent.


W1/Q1. Matched spatial propagation. Pooling uses the same self-inclusive 0.06D neighborhood and is applied to pre-sigmoid composite log scores.

ArmAUROC_occR@1
η-weighted LLR0.818 / 0.87263.95 / 71.77
Same, mean-smoothed0.902 / 0.92771.09 / 70.41
Same, max-pooled0.894 / 0.92061.90 / 70.41
Two-component with ρ replacing η0.863 / 0.92666.55 / 73.92
Type-aware, submitted0.868 / 0.92767.23 / 74.04
Two-component, mean-smoothed0.902 / 0.92370.86 / 68.82
Type-aware, mean-smoothed0.918 / 0.93972.34 / 74.26

The reviewer is correct that per-track baselines were insufficient. Smoothed LLR exceeds our submitted model under SD by +0.034 [+0.029, +0.039] AUROC and +3.9 pp [+1.1, +6.7] R@1. Under SigLIP, ranking is unresolved (+0.000 [−0.003, +0.003]); R@1 favors our model by 3.6 pp [1.4, 5.9]. Max pooling can induce exact top-score ties; R@1 retains the original first-index tie rule.

After seeing these results, we applied the same mean smoothing to our scores. Our model then exceeds smoothed LLR by +0.016 [+0.012, +0.020] / +0.012 [+0.009, +0.016] AUROC, with R@1 gains of +1.2 pp [−0.9, +3.4] / +3.9 pp [+1.6, +6.2]. This supports an additional benefit of the complete model under matched score smoothing; it does not isolate the partial-density term from the gate and other changed components. The two-component model using ρ is also not an isolated partial-density ablation. We will add these stronger comparators and narrow the attribution.


Q2. Two-component selection. R@1 is 64.97 / 71.77 for ordinary censoring and 67.23 / 74.04 for our model. The complete refinement adds +2.27 pp [−0.45, +4.86] / +2.27 pp [+0.45, +4.10]; the SD interval includes zero. Under SigLIP, ordinary censoring already gains 14.29 pp over mean-all, explaining most of Table 2's 16.55-point gain. This is a complete-refinement contrast, not an isolated partial-channel effect.


Q3. Retrained MLP ablation with matched gates. We use App. D's core population (120 scenes/672 episodes), 48 development scenes and 10 training seeds. App. D's oracle-grouped ρ uses ground-truth instance groups and estimated validity. The new spatial rows use the submitted γ=0.06 gate during both training and evaluation. Entries are mean±SD across seeds.

ModelSDSigLIP
Mixture, grouped0.8710.932
MLP, grouped0.876±0.0050.934±0.001
Mixture, spatial0.8720.926
MLP, spatial0.877±0.0030.929±0.001
MLP without ρ0.872±0.0050.923±0.002
MLP without η and ρ0.873±0.0040.914±0.001

Spatial MLP minus ρ-absent MLP is +0.0045±0.0033 / +0.0056±0.0029 (mean±SD of paired seed differences). Scene-bootstrap intervals for the seed-averaged differences are [+0.002, +0.007] / [+0.003, +0.008], conditional on these ten seeds.

Spatial MLP minus spatial mixture is +0.0051±0.0030 / +0.0032±0.0012, with intervals [+0.000, +0.010] / [+0.001, +0.006]; the SD lower endpoint rounds to zero. Thus the grouping-free feature helps the learned model in this setting, whose mean performance is slightly higher than the structured mixture. This supports the feature's utility without establishing necessity of the three-component form.

The SD model without η and ρ reaches 0.873, but retains R, β and the frustum flag F; this point estimate does not establish equivalence or make validity generally unnecessary. The earlier low-data advantage concerns grouped ρ and only 3/5 successful one-scene fits. The new full-data spatial run does not establish a spatial low-data advantage.


W2/Q4. External validity. We have not completed ScanRefer/Nr3D evaluation. Nr3D supplies object candidates, so a new proposal detector is not intrinsically required. The missing work is adapting our multi-view track pipeline and validating track-to-object scoring and evaluation. Surface-visibility annotations enable our diagnostic metrics but are not prerequisites for object-level evaluation.


The primary confirmatory fusion gains concern analytic support; real-data transfer is readout-dependent, Replica ranking gains remain unresolved, and the estimated-geometry experiments are limited by surface recovery. We will make the surface-ranking and target-track-selection scope explicit and avoid claiming general object-localization gains.

AuthorsOpenReview ↗

Follow-up on our author response

We thank the reviewer again for the constructive feedback. As the discussion period approaches its end, we would like to kindly draw your attention to our author response, which includes new experiments directly addressing your main concerns.

Specifically, we added matched spatial-propagation baselines, evaluated the two-component mixture on R@1-track, and retrained the MLP with and without the spatial visibility feature. These results help better characterize the contributions of spatial coupling and the proposed observation model. We also acknowledge the remaining limitations.

If time permits, we would greatly appreciate your feedback on whether these additional results address your concerns.

Thank you again for your time and consideration.

Thank you again for running the matched baselines and for reporting them openly.

To state my remaining concern plainly: the new results show that most of the headline gain comes from spatial propagation and not from the three-component likelihood. Mean-smoothed η-weighted LLR matches the submitted model under SigLIP and exceeds it by 0.034 under SD. The two-component model with ρ in place of η recovers nearly all of the improvement. What remains for the full model under matched smoothing is +0.016 / +0.012, and that variant was defined after looking at the sealed partition, so it is post hoc.

That said, the residual gain is consistent across both backbones, the evaluation is trustworthy, and the benchmark and the partial versus full occlusion observation are useful on their own. I am raising my rating to 3 (Weak acceptance) on the understanding that the revision will do the following.

  1. Report the smoothed LLR and ρ-based two-component baselines in the main results, next to the original comparators.
  2. State clearly that the smoothed type-aware model is a post hoc variant and that the preregistered gains were measured against baselines without spatial coupling.
  3. Reframe the contribution so that the neighborhood visibility signal is credited for most of the gain and the three-component form for the smaller remainder.
  4. Put the SigLIP R@1 gain in context, since ordinary censoring already obtains 14.29 of the 16.55 points.
  5. Limit the claims to surface ranking and target-track selection on analytic support, as you proposed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.