acceptodds
Under review as a conference paper at ICLR 2027

Where Does the Uncertainty Come From? Four-Channel Information Decomposition for Active 3D Capture

Abstract

Active 3D reconstruction usually picks the next viewpoint by a scalar expected information gain, which says how much a view teaches but not which part of the scene it constrains or whether that information belongs to the persistent scene. We introduce PRISM, a task-aware next-best-view method for Gaussian Splatting that decomposes a candidate view's persistent information gain into geometry and opacity, base appearance, and view-dependent appearance. Under a local linear-Gaussian approximation, a Shapley allocation yields non-negative attributions that sum exactly to the nuisance-marginalized gain, while transient content enters as a separate conditional-information term that cannot inflate the persistent shares. Task weights recombine the attributions for different objectives over one candidate pool, and the same predictive covariances turn the residuals of arriving images into nuisance evidence for later rounds. On surface reconstruction, low-light capture, and reflective scenes, each task's utility beats total-information, Fisher-information, and mutual-information planners on its own metric, most inside the regions the task cares about; on a fixed drone route it slows where its evidence lies; and in a cluttered scene, nuisance-aware scoring cuts the share of selected rays on moving objects from 0.31 to 0.12.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Active Reconstruction, Next Best View, Uncertainty

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_c1bhOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper decomposes the information gain of novel view into three components: geometry, base color, and view-dependent color (G,B,V) using a Shapley method and re-weights each component for specific tasks. It segments the transient objects by comparing the rendering residual against the persistent uncertainty and can segment out the transient model.

Critical strengths
  1. Using rendering error and persistent uncertainty to segment out transient objects is a novel idea.
  2. Decomposing and re-weighting the information from geometry, base color, and view-dependent color shows better performance for different tasks.
Critical weaknesses
  1. It seems that the task weight α\alpha is manually tuned. How sensitive is the result w.r.t. alpha?
  2. I am still concerned about the experiment setup. I think a trivial solution to decompose the information is to only use the variance matrix for G, B, and V to compute the information gain, without the shapely method. For example, for FisherRF in 2DGS scene tasks, one can only use the covariance matrix for G to compute the information gain to select views, ignoring the other parameters. I wonder if PRISM could even beat these baselines.
Questions for authors
  1. What is Π(P)\Pi(P) in equation 5 ?
  2. What are the values of alpha_G, alpha_B, alpha_V?
  3. Since you cite RobustNeRF, which also uses residuals to segment out transient objects. Could you compare the IoU of masks for transient objects to RobustNeRF?
2 comments

Your trivial decomposition is the right control for what the allocation adds, so we answer it first.

W2. We tested this trivial decomposition directly on PRISM's posterior: computing the log-determinant information gain F_v({b}; R) from only the target block's covariance (the other blocks held as known, with no Shapley averaging) under identical one-hot weights. Across all three static testbeds (Table 6, submitted Table 5) and all three tasks of the new dynamic closed-loop test with real distractors (Section 4.5, Table 2), PRISM outperforms single-block scoring:

  • Static 2DGS SfM depth error (x10^-3, lower better): single-block 4.88 vs. PRISM 4.79 (unweighted total information: 4.85).
  • Static low-light PSNR (dB): single-block 39.02 vs. PRISM 39.36 (total information: 39.06; note PRISM 39.36 and FisherRF 39.31 dB are level in low light).
  • Static reflective-pixel PSNR (dB): single-block 17.38 vs. PRISM 17.51 (total information: 17.44).
  • Dynamic RobustNeRF closed loop (android, crab2, yoda; 3 seeds, 9 paired runs per task, beta = 0, online Psi; Tasks B and V measured against a reference model's rendering): Task G SfM depth error (x10^-3) single-block + Psi 30.2 vs. PRISM 28.0 (PRISM wins 7/9 runs); Task B base-color PSNR 19.6 vs. 20.0 dB (8/9); Task V specular PSNR 16.5 vs. 16.9 dB (8/9).

On all three static testbeds, single-block scoring is no better than unweighted total information (4.88 vs. 4.85, 39.02 vs. 39.06 dB, 17.38 vs. 17.44 dB), consistent with standalone gains treating the other uncertain blocks as known and double-counting shared pixel evidence (summing to 1.5x I_pers on 3DGS; Figure 7c). Rescaled singletons—normalizing standalone gains to sum to I_pers—also improve on total information (Table 6: 4.83, 39.09 dB, 17.46 dB). PRISM is ahead of them on every static column (4.79, 39.36 dB, 17.51 dB) and in 7 or 8 of 9 dynamic runs (Table 2), though the static margins over rescaled singletons are small.

Running single-block scoring on PRISM's own posterior, Jacobians, and prior isolates the allocation rule cleanly, whereas restricting FisherRF to one parameter block would simultaneously change the uncertainty model (a diagonal Hessian approximation instead of a block log-determinant). A block-restricted FisherRF baseline was not evaluated; this is stated as a limitation in Section 5.

Q1. In Equation (5), P = {G, B, V}, Pi(P) is the set of the six permutations (orderings) of P, and A_b^pi is the set of blocks preceding b in ordering pi. That was our omission, not only a rendering issue: Pi(P) was not defined in the submitted text and is now defined immediately below Equation (5) in Section 3.2.

W1 / Q2. The impression that alpha was tuned came from inaccurate wording in our submitted Reproducibility Statement, which listed alpha among the quantities fixed on validation data. In all task experiments, alpha is simply the un-tuned one-hot vector matching the task definition: alpha^(G) = (1, 0, 0), alpha^(B) = (0, 1, 0), and alpha^(V) = (0, 0, 1), with uniform weights (1/3, 1/3, 1/3) as the total-information control (now written explicitly in Section 3.3, and corrected in the Reproducibility Statement; only the evidence parameters tau, pi_T and the gate constants were fixed on validation data). For sensitivity, the paper reports the uniform and one-hot endpoints on all testbeds (Table 1, including off-diagonal mismatched weights) and one-hot weights under nuisance (Table 2); intermediate alpha values were not run.

Q3. No dataset-level transient mask IoU and no comparison against RobustNeRF's masks were computed. In Figure 3 (bottom row), one selected clutter frame prints IoU 0.65 (with 1% of static pixels flagged) under the persistent-covariance reference c_r versus IoU 0.39 (10% static flagged) under noise only, which is a single-frame illustration rather than a segmentation benchmark. Across probe frames at matched 0.80 transient recall (Section 4.2, Table 4), persistent-region false flags drop from 0.40 (noise only) to 0.39 (trace-matched c_r) and 0.35 (full 3x3 covariance). Here residual normalization supplies Psi for next-view scoring before fitting rather than training masks: in the reported runs Psi enters the selection score only, the posterior is accumulated with sigma_0^2 (not nuisance-aware), and images are trained on unmodified. Removing residual normalization in Table 2 degrades means across Tasks G (29.5 vs. 28.0, 7/9), B (19.7 vs. 20.0 dB, 7/9), and V (16.6 vs. 16.9 dB, 6/9).

Summary of changes: single-block vs. PRISM static (Table 6) and dynamic (Section 4.5, Table 2) comparisons detailed; Pi(P) defined under Eq. (5); one-hot alpha clarified in Section 3.3 and Reproducibility Statement. Thank you for the baseline suggestion; it sharpened what the allocation does and does not buy.

The author's response resolved most of my concerns. Comparison with the single-block baseline shows that the Shapely method is effective. I will maintain my score and wish the best for the authors.

Reviewer_A8YPOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PRISM, a next-best-view method for Gaussian Splatting. Instead of ranking candidate views by one scalar information gain, it splits that gain by its source, such as geometry and appearance. Interestingly, the experiments in this paper cover many aspects to evaluate the reconstruction quality, like geometry (2DGS on Ref-NeRF scenes and DTU scan 24), base appearance under low light (LOM), view-dependent appearance (reflective Ref-NeRF scenes), a cluttered scene with distractors, and speed of planning along a fixed drone route in a UE5 city scene.

Critical strengths
  1. The motivation is clear and interesting. Decompositing uncertainty value to find the source of information is a novel perspective. Two views with the same total information gain can be valuable for very different reasons, and a scalar gain hides that. Fig. 1 makes this point nicely, and connecting the choice of view to the downstream task is a sensible direction for active reconstruction.

  2. Various reconstruction metrics are tested to verify the overall quality of newly selected views. The author consider cases including view sythesis, noisy image based reconstruction and view-dependent scenes, which are plentiful.

  3. In the presented results, both reconstruction quality metrics and path efficiency metrics are verified improvements, demonstrating the practical usefulness of active reconstruction.

Critical weaknesses
  1. Some related works in 3DGS active reconstruction could be discussed thoroughly:

[1] Li, et al. "Active Learning of 3D Gaussian Splatting with Consistent Region Partition and Robust Pose Estimation." ICLR, 2026.

[2] Chen, et al. "Activegamer: Active gaussian mapping through efficient rendering." CVPR, 2025.

  1. Would the authors consider revising the subtitles of section 3.3 and 3.5 to describe the method accurately?

  2. Additional ablation on the task weights can strengthen the soundness. The main task experiments use one-hot weights, with uniform weights as the total-information control. An ablation on the weights alpha value and report the task metrics and selection changes can help in understanding the method. This would show whether the method gives useful trade-offs between geometry, base appearance, and view-dependent appearance.

Questions for authors

Please see the weaknesses.

1 comment

Thank you specifically for pointing us to Li & Cheung (ICLR 2026) and ActiveGAMER (CVPR 2025), and for asking how selection and task metrics behave across task weights alpha. Your review centers on related-work positioning (W1), descriptive section headings (W2), and alpha trade-offs (W3).

W3. Because the persistent attribution vector S(v) = (S_G, S_B, S_V) is computed once per candidate view and alpha enters linearly in U_alpha(v) = alpha^T S(v) - beta S_T(v) (Eq. 6), changing alpha re-ranks the same candidate pool without recomputing coalition log-determinants. The revised paper reports the two informative endpoints of the simplex on every testbed—uniform weights (1/3, 1/3, 1/3), which reproduce total-information ranking I_pers, and the one-hot task vectors alpha^(G), alpha^(B), alpha^(V)—as well as one-hot weights with online Psi in the new joint closed-loop experiment (Section 4.5, Table 2); closed-loop sweeps over intermediate interior alpha values were not run.

Across the static testbeds (Tables 1 and 6), moving from uniform weights to the task-matched one-hot vector improves every target metric, whereas singleton block scoring without Shapley does not:

  • 2DGS SfM depth error (x10^-3, lower better): uniform 4.85 -> PRISM alpha^(G) 4.79 (vs. 4.88 for singleton scores).
  • Low-light PSNR: uniform 39.06 dB -> PRISM alpha^(B) 39.36 dB (vs. 39.02 dB for singleton; level with FisherRF at 39.31 dB).
  • Reflective-pixel PSNR: uniform 17.44 dB -> PRISM alpha^(V) 17.51 dB (vs. 17.38 dB for singleton; +0.19 dB and 8/9 paired runs vs. GauSS-MI at 17.32 dB; Appendix D).

The off-diagonal entries of Table 1 show the trade-off across G, B, and V at the simplex vertices: each one-hot utility leads on its matched metric and gives up performance on mismatched ones—for example, in low light alpha^(B) reaches 39.36 dB while alpha^(V) drops to 38.83 dB; on reflective scenes alpha^(V) reaches 17.51 dB on reflective pixels (21.62 dB non-reflective) while alpha^(B) reaches 17.29 dB (21.68 dB non-reflective) and uniform weights reach 17.44 dB (21.77 dB non-reflective). In the new dynamic closed loop on three RobustNeRF distractor scenes (Section 4.5, Table 2; 9 paired runs/cell), one-hot PRISM also achieves the best deployable mean on Task G (28.0 vs. 29.9 symmetric and 30.2 single-block + Psi), Task B (20.0 vs. 19.6 dB), and Task V (16.9 vs. 16.5 dB).

On implementation scope (now stated in Section 4 Setup and Appendix C): the mini-3DGS backend uses 12,000 to 18,000 isotropic primitives without densification and parameterizes V as a 3x3 matrix of direction coefficients, while the 2DGS backend runs at one-eighth image resolution with degree-1 SH in V. A full anisotropic 3DGS backend with densification and degree-3 SH was not tested.

W1. Section 2 and Table 7 (Appendix D) now discuss both Li & Cheung (ICLR 2026) and ActiveGAMER (Chen et al., CVPR 2025), alongside OUGS, GO-PRE, Zhu et al., and DynActiveGS. They differ from PRISM along two axes:

  • Spatial/region partitioning and pose robustness vs. parameter-role attribution: Li & Cheung partition a 3DGS scene into consistent regions using Gaussian attributes and visibility, score semantic feature variance, and jointly optimize noisy camera poses (similarly, OUGS partitions spatially by object mask). PRISM instead decomposes information gain across rendering parameter roles (G, B, V) and assumes known camera poses: under pose error, alignment residuals would exceed c_r and enter q_k and Psi as if they were transient.
  • Exploration criteria: ActiveGAMER is an RGB-D exploration and mapping system driven by a rendering-based geometric and photometric information gain, rather than a block-posterior source decomposition over a shared candidate pool. Neither ActiveGAMER nor Li & Cheung was evaluated experimentally here, and pose uncertainty is not modeled; both points are noted in Section 5.

W2. Thank you, this is fixed. Section 3 now uses accurate descriptive subtitles throughout:

  • 3.1 Persistent Parameter Blocks and the Predictive Model
  • 3.2 Shapley Allocation of Persistent Information Gain
  • 3.3 Task-Weighted Selection Utility
  • 3.4 Nested Allocation of Transient Nuisance
  • 3.5 Residual Normalization and Nuisance Evidence at Arrival (And in Section 4.3 and Figure 4, the scalar baseline on the UE5 route is clarified as the unmarginalized joint information I_aug.)

Summary of changes: Li & Cheung and ActiveGAMER discussed in Section 2, Section 5, and Table 7; Section 3 subtitles revised; uniform vs. one-hot trade-offs clarified with dynamic closed-loop results in Section 4.5 (Table 2) and backend scope disclosed in Section 4.

Reviewer_bqBaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a task-aware next-best-view framework for Gaussian Splatting that decomposes expected information gain into geometry (G), base appearance (B), and view-dependent appearance (V). Under a local linear-Gaussian approximation, Shapley attribution distributes the nuisance-marginalized information gain among these sources, which are then weighted according to the target acquisition task. Transient content (T) is handled separately through nuisance-precedence allocation, supported by theoretical analyses of its properties. The method also estimates nuisance evidence using covariance-normalized residuals and incorporates it into subsequent viewpoint selection. Experiments on geometry reconstruction, low-light and reflective scenes, transient clutter, and a simulated urban route report improvements over existing information-based acquisition baselines.

Critical strengths

1. Meaningful problem formulation. The paper formulates the problem of allocating a common information gain among interacting rendering sources without double-counting. By connecting source-level information attribution to task-aware view selection, the proposed framework enables different reconstruction objectives to be considered within a unified information-gain formulation.

2. Mathematically grounded analysis of nuisance-aware attribution. The paper provides a theoretical analysis of the proposed nuisance-precedence allocation. In particular, Theorem 2 quantifies the over-attribution of persistent information under symmetric allocation, while Theorem 3 establishes the uniqueness of the proposed allocation under specified axioms and constraints. These results provide a mathematical justification for the asymmetric treatment of nuisance information.

3. Practical integration of nuisance evidence into the acquisition loop. The paper uses predictive covariance to normalize residuals and estimate nuisance evidence from newly acquired observations. By retaining this evidence on Gaussian primitives and incorporating it into subsequent viewpoint selection, the proposed method provides a practical mechanism for handling transient content during active acquisition.

Critical weaknesses

1. Presentation clarity. I found the paper considerably difficult to follow throughout. Several technical terms, such as "persistent shares," "nested allocation," "reference scale," and "trace-matched scale," are introduced without sufficient context, particularly before readers are familiar with the proposed method. The motivation and significance of the theoretical contributions, including the uniqueness and quantitative bounds of the proposed allocation, are also difficult to understand when first introduced. Even in the method and experimental sections, the relationships between individual components and their roles in the overall framework are not always clearly explained. I believe the presentation needs substantial improvement to make the main ideas and contributions accessible to readers.

2. The performance improvements are relatively modest. Performance improvements over the baselines are modest in several experiments. For example, Table 1 reports improvements from 4.99 to 4.79 in SfM depth error (×10⁻³), 39.31 to 39.36 dB in low-light PSNR, and 17.22 to 17.51 dB in reflective-pixel PSNR compared with FisherRF. Although larger improvements are reported in the task-specific focus regions, it is still unclear whether these gains are sufficiently meaningful. Furthermore, the qualitative results in Figure 3 do not show clearly distinguishable improvements over the baselines. Additional quantitative evidence supporting the significance of these improvements would strengthen the evaluation.

3. Insufficient comparison and differentiation from related methods. The main evaluation in Table 1 primarily compares against baselines that do not explicitly consider source-level attribution or task-specific objectives, while Table 5 mainly provides internal allocation ablations. The paper would benefit from comparisons with related methods, particularly those considering source-specific uncertainty (e.g., OUGS) and task-specific information gain (e.g., GO-PRE). In addition, the authors should more clearly distinguish their approach from existing methods that address source-specific uncertainty (e.g., [1]), task-specific information gain (e.g., [2]), or transient/dynamic content during acquisition (e.g., [3]). Including relevant baselines where applicable, along with a more detailed discussion of these differences, would help establish the contributions of the proposed method.

4. Insufficient experimental coverage. Although the paper evaluates several tasks, the number of test scenes for each task is relatively small (Table 4), particularly for low-light and reflective scenes. The authors should consider expanding the evaluation to more diverse environments or justify whether the current scale is sufficient to demonstrate generalizability across environments. Furthermore, the task-aware experiments exclude nuisance uncertainty, while the nuisance-aware acquisition experiment uses uniform task weights and only one scene. Therefore, quantitative evaluation of the combined effectiveness of task-aware attribution and nuisance handling in closed-loop acquisition remains insufficient.

5. Insufficient detailed analysis. Section 4.4 appears intended to provide a detailed analysis of the proposed method, but its purpose is not sufficiently clear. The section seems to conflate two different questions: (i) whether the proposed parameter categorization appropriately captures the information sources relevant to the target evaluation scenarios, and (ii) whether the source-level attributions are accurately computed despite the approximations involved. The authors should clarify which aspects the analysis is intended to validate and explain the importance of each.

6. Novelty should be stated more clearly. The paper should more clearly articulate its novel contributions. In particular, the source-level information attribution applies the existing Shapley-value framework to relatively straightforward categories of 3DGS parameters, making the novelty of this component appear limited. The authors should better clarify what distinguishes the proposed method from existing approaches and where its main novelty lies.

<references>

[1] Li, Haiyi, et al. "OUGS: Active View Selection via Object‐aware Uncertainty Estimation in 3DGS." Computer Graphics Forum. 2026.

[2] Zhu, Jingsen, Silvia Sellán, and Alexander Terenin. "A Bayesian Approach for Task-Specific Next-Best-View Selection with Uncertain Geometry." arXiv preprint arXiv:2605.05095 (2026).

[3] Duan, Hongbo, et al. "DynActiveGS: Active Gaussian Splatting for Dynamic Scene Reconstruction." arXiv preprint arXiv:2608.01178 (2026).

Questions for authors
  1. Given the relatively modest improvements in Table 1 and the limited number of test scenes, can the authors provide evidence that the gains are statistically reliable and generalizable across different environments?

  2. Beyond applying Shapley attribution to predefined 3DGS parameter groups, what is the main methodological novelty of PRISM compared with existing source-aware and task-specific view selection methods?

  3. Theorem 3 establishes the uniqueness of the proposed allocation under specific axioms and nuisance-precedence constraints. Could the authors further justify why these assumptions are necessary or desirable for active acquisition? In particular, what is the practical significance of this uniqueness result, given that the experimental comparisons do not directly evaluate alternative nuisance-aware allocations?

3 comments

Part 1 covers novelty (W6/Q2), Theorem 3 (Q3) and related work (W3); Part 2 covers the evidence (W2/W4/Q1), Section 4.4 (W5) and presentation (W1).

W6 / Q2. Applying Shapley values to three parameter groups would be a small step on its own; we agree, and the submission did not state the contribution sharply enough. Conservation and uniqueness alone do not prove better acquisition. The methodological contribution lies in preventing two specific mis-allocations of shared pixel evidence within one predictive model, and reusing that model's covariances at arrival:

  1. Avoiding double-counting across interacting persistent blocks: G, B, and V are independent in the block posterior, yet their effects meet in the same pixels through the rendering weights. Scoring one block alone F_v({b}) treats the other uncertain blocks as known and double-counts shared evidence, summing to 1.5x I_pers on 3DGS scenes and 1.09x on 2DGS (Section 4.4, Figure 7c). Under identical one-hot weights, singleton scores trail PRISM—and even unweighted total information—on all three static tasks (Table 6), and Single-block + Psi trails PRISM in 7/9, 8/9, and 8/9 runs in the dynamic closed loop (Table 2).
  2. Preventing transient variance from inflating persistent credit: treating T as a symmetric fourth Shapley player evaluates persistent marginals with unobserved disturbance held known, over-crediting persistent blocks by [Delta_v/4, 3 Delta_v/4] (Theorem 2, Appendix A.3). Keeping T in the effective noise R_T = R_0 + Psi and separating S_T = I(xi; Y_v | theta) preserves I_pers; in Table 2 (beta = 0), PRISM beats symmetric four-way allocation in 8/9, 6/9, and 8/9 runs.
  3. The same block covariances set the residual scale c_r on arrival (Eq. 8), lowering persistent-pixel false flags at matched transient recall from 0.40 (noise only) to 0.39 (trace-matched) and 0.35 (full 3x3 covariance; Section 4.2).

Q3. Each condition in Theorem 3 (Section 3.4, Appendix A.5) has an operational meaning:

  • Efficiency: the persistent shares sum to the persistent information gained.
  • Symmetry among persistent blocks: the shares do not depend on an arbitrary inclusion order. A fixed inclusion order shifts shares by a median of 16 percentage points and changes the dominant source in more than half of the candidates (Section 4.4, Figure 7d).
  • Real-linearity: the allocation of a sum of games is the sum of the allocations, so changing units (nats or bits) scales all shares alike.
  • Feasible null player: a block that the view cannot inform gets zero.
  • Nuisance precedence is a modeling premise that restricts the game domain (every nonempty feasible coalition contains T), not an axiom. It matches acquisition because the disturbance realization in the next image is unobserved when the view is scored, so no persistent marginal is evaluated with T held known. Uniqueness holds only under this premise: on the full four-player game space, the standard Shapley axioms give the symmetric four-player value instead (Appendix A.5 contrasts the two).

Practical meaning: Theorem 3 shows that once nuisance precedence is adopted, the nested rule involves no further arbitrary weighting; empirical impact against an alternative nuisance-aware rule—which the submission lacked—is now tested directly via the symmetric four-way planner in Table 2.

W3. The submission listed OUGS only inside a group citation without discussion, so your criticism is fair. Section 2 and Table 7 (Appendix D) now discuss OUGS, GO-PRE (Song et al., arXiv 2607.29037), Zhu, Sellán and Terenin (arXiv 2605.05095), DynActiveGS (arXiv 2608.01178), ActiveGAMER and Li & Cheung. OUGS partitions uncertainty spatially by object, whereas PRISM attributes it by rendering role; GO-PRE and Zhu et al. optimize a target-view or downstream-task objective directly, whereas PRISM reweights parameter-role information without modeling task-loss reduction; DynActiveGS models motion-induced uncertainty and down-weights it in optimization, whereas PRISM marginalizes a residual-based Psi in scoring only. None of these methods was run experimentally; Section 5 states this limitation.

W2 / W4 / Q1. Full-frame gains in Table 1 are modest and scene coverage is small (Table 5, submitted Table 4): 5 scenes in low light (15 paired runs), 3 in reflective 3DGS (9 paired runs), 3 in 2DGS surface (unequal run counts per scene) plus DTU scan 24, 1 clutter scene in Section 4.2, and 3 RobustNeRF scenes in Section 4.5. Conclusions are explicitly bounded to these tested scenes (Section 5).

On static testbeds (Tables 1 and 6):

  • 2DGS SfM depth error (x10^-3): PRISM alpha^(G) reaches 4.79 vs. 4.99 (FisherRF), 4.82 (GauSS-MI), 4.85 (total information), and 4.88 (singleton).
  • Low-light PSNR: PRISM alpha^(B) (39.36 dB) and FisherRF (39.31 dB) are level (+0.05 dB), ahead of total information (39.06 dB) and singleton (39.02 dB).
  • Reflective PSNR: PRISM alpha^(V) reaches 17.51 dB vs. 17.22 dB (FisherRF), 17.32 dB (GauSS-MI; PRISM wins 8/9 paired runs, and 7/9 vs. random), 17.44 dB (total information), and 17.38 dB (singleton).
  • Focus regions (defined from reference models/images only, never used in selection or fitting; Appendix D): +0.6 dB over total information on the darkest fifth of low-light frames, +1.1 dB over total information (+1.4 dB over FisherRF) on glossy segments, and 2.94 mm vs. 3.39 mm Chamfer distance on DTU scan 24. Figure 3 displays selected largest-margin frames, as its caption notes.

To test non-uniform task weights and online Psi together in one closed loop across multiple scenes, Section 4.5 (Table 2) evaluates three RobustNeRF distractor scenes (android, crab2, yoda; 3 seeds, 9 paired runs/cell, one-hot alpha, beta = 0; tau and pi_T fixed once on clean yoda; Psi enters selection only while the posterior uses sigma_0^2). Task G is SfM median relative depth error (x10^-3, lower better); Tasks B and V are base-color and top-20% specular PSNR (dB) against a reference model's rendering:

Planner (9 paired runs/cell; PRISM wins in brackets)Task G (x10^-3)Task B (dB)Task V (dB)
PRISM28.020.016.9
Single-block + Psi30.2 (7/9)19.6 (8/9)16.5 (8/9)
Rescaled + Psi29.8 (7/9)19.7 (7/9)16.6 (8/9)
Symmetric four-way29.9 (8/9)19.6 (6/9)16.5 (8/9)
No residual norm.29.5 (7/9)19.7 (7/9)16.6 (6/9)
Oracle (GT mask)27.6 (3/9)20.1 (3/9)17.0 (2/9)

PRISM has the best deployable mean on all three tasks (the ground-truth mask oracle is higher; Task B vs. symmetric four-way and Task V vs. no residual normalization are 6/9).

W5. Section 4.4 is now split into two paragraphs with distinct questions:

  1. Does the G/B/V parameter grouping track the underlying scene source? (Dominant share matches the scene's stressed source in 84%, 79%, and 88% of candidates; V share correlates with specular fraction and G share with unseen primitives in Figure 7a,b; single-source deficit rank correlations are 0.33-0.71 for G and 0.76-0.88 for V.)
  2. Does the Shapley allocation matter beyond simpler block rules, and how sensitive is the model approximation? (Standalone block scores sum to 1.5x I_pers on 3DGS; Shapley beats singleton, rescaled, fixed-order, and focused rules in Table 6; Appendix D checks the block-diagonal approximation itself: restoring within-primitive cross-block terms shifts the joint information by 2.5% on average.)

W1. Sections 3.1-3.5 now use descriptive headings, motivate Theorems 2 and 3 in plain prose before citing formal statements, end Section 3.5 with a five-step "One round" walkthrough connecting scoring, arrival, and fitting, clarify Section 4.3's scalar baseline as joint information I_aug, and define the four terms at first use:

  • Persistent share (Section 3.2): the part S_b(v) of the nuisance-marginalized persistent gain I_pers(v) attributed to persistent block b, with S_G + S_B + S_V = I_pers(v).
  • Nested allocation (Section 3.4): holding transient variance Psi in the effective noise while G, B, and V split I_pers(v) by the Shapley value, and assigning the remaining conditional information S_T(v) = I(xi; Y_v | theta) to T so all four sum to I_aug(v).
  • Reference scale (Section 3.5): the per-channel residual variance c_r predicted by the persistent model before fitting, equal to sigma_0^2 plus one third of the trace of the summed block predictive covariances (Eq. 8).
  • Trace-matched scale (Section 3.5): using one third of the trace of the 3x3 per-ray predictive covariance (an isotropic per-channel variance) in place of the full 3x3 covariance matrix.

Summary of changes: novelty and the role of the Theorem 3 premises stated more sharply (Sections 1, 3.4); OUGS, GO-PRE, Zhu et al., and DynActiveGS positioned in Section 2 and Table 7; 3-scene joint closed-loop test added in Section 4.5 (Table 2); Section 4.4 split in two and Section 3 rewritten with inline definitions and a round walkthrough.

Reviewer_bqBaOpenReview ↗

Official Comment by Reviewer bqBa

The following major concerns remain unresolved, so I keep my rating.

  • Presentation clarity (definitions added in Section 3, but the Introduction is largely unchanged; still hard to follow)
  • Comparison with the methods I raised (no experimental comparison; justification for omitting it is not convincing)
  • Modest gains (significance, robustness, and generalizability not convincingly demonstrated)
  • Number of test environments (static testbeds unchanged; joint test limited to three scenes)
Reviewer_TaW7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a next-best-view method for Gaussian Splatting. It uses Shapley value to decompose persistent information gain into geometry G, base appearance B and view-dependent appearance V. It also introduces an asymmetric channel to distinguish transient content from persistent ones. The authors conduct experiments on surface reconstruction, low-light cases, reflective scenes and shows overall gains over information-based baselines.

Critical strengths
  1. The method is principled and has a clean theoretical formulation. The decomposition into G, B, V sums to the persistent gain, with each term non-negative, and has explicitly considered the interactions between rendering components. The asymmetric treatment of nuisance information is intuitive and well-motivated.

  2. Thorough disgnostic experiments of the attribution. The paper evaluates different reconstruction objectives and careful comparisons with alternative strategies, as shown in Table 1&5. I also find the limitation section helpful on explaining the underlying assumptions of the method and where it may not hold.

Critical weaknesses
  1. The overall performance gain is relatively small (compared to simpler alternatives) and without uncertainty measures. According to Table 1, this comparison is around (4.85->4.79, 39.06->39.36) dB. Can the authors provide uncertainty estimates or statistical significance or standard deviations of those values?

  2. The introduction writing can be clearer by stating the problem first, then motivation and then technical details. Currently it mixes the authors' own observations with related work. Each observation can then directly lead to one of the three method contributions, instead of scattering in difference places.

Questions for authors

Q1. Please see weakness 1.

Q2. How much of the performance gain is from Shapley allocation vs. task-specific source weighting?

Q3. It seems that task weighting and nuisance handling are never evaluated together: task experiments has Ψ=0\Psi = 0 while the one with transient objects (Sec. 4.2) uses uniform weights. Can the author provide additional experiments results using closed-loop experiment with moving objects?

3 comments

Your review focuses on whether the modest full-frame gains hold up with run-level dispersion accounted for, how much comes from Shapley allocation versus task weighting, whether task weighting and nuisance handling work together in a closed loop, and the structure of the Introduction.

W1 / Q1. Agreed that the full-frame mean differences in Table 1 are small. Because seeds within a scene share the scene geometry and candidate pool, runs within a scene are not independent samples; we report means, differences of means, and paired win counts where we list them. Scale and pairing structure (Table 5, Appendix D): low light has 5 scenes (buu, chair, sofa, bike, shrub) x 3 seeds = 15 paired runs; reflective 3DGS has 3 scenes (toycar, gardenspheres, sedan) x 3 seeds = 9 paired runs; 2DGS surface has the same 3 scenes averaged over the available runs per scene, which are unequal across scenes, so no paired win counts are reported for 2DGS. (Note: in the submitted Figure 8 the low-light PRISM curve came from an earlier run; that panel is removed in the revised PDF and Table 1 is unchanged.)

On full frames (Table 1), comparing PRISM against unweighted total information (the 4.85 and 39.06 baselines you noted):

  • 2DGS SfM depth error (x10^-3, lower is better): PRISM alpha^(G) reaches 4.79 vs. 4.85 for total information (-0.06), 4.82 for GauSS-MI (-0.03), and 4.99 for FisherRF (-0.20).
  • Low-light held-out PSNR: PRISM alpha^(B) reaches 39.36 dB vs. 39.06 dB for total information (+0.30 dB) and 38.36 dB for GauSS-MI (+1.00 dB). Against FisherRF (39.31 dB), the difference is +0.05 dB (and alpha^(G) reaches 39.29 dB); PRISM and FisherRF are level in low light, and the revised text states this plainly (Sections 1, 4.1, 5).
  • Reflective-pixel PSNR: PRISM alpha^(V) reaches 17.51 dB vs. 17.44 dB for total information (+0.07 dB) and 17.22 dB for FisherRF (+0.29 dB). On reflective 3DGS, PRISM alpha^(V) beats GauSS-MI (17.32 dB, +0.19 dB) in 8/9 paired runs and random selection in 7/9 (Appendix D).

Inside focus regions defined from the reference model and photographs independently of selection (Section 4.1, Appendix D), margins are larger: +0.6 dB over total information on the darkest fifth of low-light frames, +1.1 dB over total information (+1.4 dB over FisherRF) on glossy segments, and 2.94 mm Chamfer distance on DTU scan 24 vs. 3.39 mm (total information), 3.62 mm (FisherRF), and 3.53 mm (GauSS-MI). Conclusions remain limited to these 3 to 5 scenes per setting (Section 5).

Q2. Table 6 (submitted Table 5, Appendix D; discussed in Section 4.4) isolates task weighting from Shapley allocation on the exact same posterior. Under uniform weights, PRISM equals total information. Applying one-hot task weights to each block's standalone gain F_v({b}) without Shapley gives the singleton row:

  • 2DGS depth error (x10^-3, lower better): total information 4.85 -> singleton 4.88 -> rescaled singletons 4.83 -> PRISM 4.79.
  • Low-light PSNR (dB): total information 39.06 -> singleton 39.02 -> rescaled singletons 39.09 -> PRISM 39.36.
  • Reflective PSNR (dB): total information 17.44 -> singleton 17.38 -> rescaled singletons 17.46 -> PRISM 17.51.

Scoring the task block alone is no better than unweighted total information on all three testbeds (4.88 vs. 4.85, 39.02 vs. 39.06, 17.38 vs. 17.44), consistent with standalone gains double-counting shared pixel evidence (summing to 1.5x I_pers on 3DGS; Figure 7c). The improvement over total information appears only when task weights are paired with a conserved allocation, with rescaled singletons recovering part of the difference and Shapley reaching the best mean on each testbed.

Q3. Agreed that the submission evaluated non-uniform weights and online Psi only separately. Section 4.5 (Table 2) runs one-hot task weights together with online Psi in a closed loop on three RobustNeRF scenes with real distractors (android, crab2, yoda; 3 seeds, 9 paired runs per cell, beta = 0; tau, pi_T fixed once on clean yoda; Psi enters candidate scoring while posterior accumulation uses fixed sigma_0^2). Task G is SfM median relative depth error (x10^-3, lower better); Tasks B and V are base-color PSNR and top-20% specular PSNR (dB) against a reference model's rendering:

Planner (9 paired runs/cell; PRISM wins in brackets)Task G (x10^-3)Task B (dB)Task V (dB)
PRISM28.020.016.9
Single-block + Psi30.2 (7/9)19.6 (8/9)16.5 (8/9)
Rescaled + Psi29.8 (7/9)19.7 (7/9)16.6 (8/9)
Symmetric four-way29.9 (8/9)19.6 (6/9)16.5 (8/9)
No residual norm.29.5 (7/9)19.7 (7/9)16.6 (6/9)
Oracle (GT mask)27.6 (3/9)20.1 (3/9)17.0 (2/9)

PRISM has the best mean among deployable planners on all three tasks (the ground-truth mask oracle is ahead on all three; Task B vs. symmetric four-way and Task V vs. no residual normalization are 6/9 splits).

W2. Section 1 now states the active-capture problem first, keeps related-method comparisons in Section 2, and links each motivating observation directly to one contribution:

  • Equal total persistent gain can constrain different interacting roles (G, B, V) -> Contribution 1: conserved Shapley attribution across G, B, V for task-weighted selection.
  • Unobserved transient content inflates joint information and symmetric persistent shares -> Contribution 2: nested nuisance allocation with bracket bounds and uniqueness.
  • Large residuals can stem from unresolved persistent structure rather than transients -> Contribution 3: covariance-normalized residual check supplying Psi and capture-time audit.

Summary of changes: low-light parity with FisherRF stated (Sections 1, 4.1, 5); joint closed-loop experiment added (Section 4.5, Table 2); Introduction aligned with the three contributions (Section 1). We hope these numbers answer your three questions; we are glad to clarify anything further.

Thanks to the authors for their response, it has addressed some of my questions. I maintain my score since the gains are relatively small and are still reported without uncertainty estimates.

Comments on the paper

3 comments

Your audit identifies four issues: a data discrepancy between Figure 8 and Table 1, the missing closed-loop test of the four-channel claim under non-uniform weights and online nuisance, the operational gap between scoring and model updates, and what the allocation adds over simpler task-specific scoring with nuisance down-weighting. We address W4/Q4 and W3/Q3 here, and W2/Q2 and W1/Q1 in Part 2.

W4 / Q4. Thank you, this is fixed. In the submitted Figure 8, the low-light PRISM curve came from an earlier run, while every other curve in Figure 8 matched Table 1. The revised Figure 8 (Appendix D) drops the low-light panel and retains the 2DGS and reflective panels; Table 1 is unchanged and reports the low-light endpoints. In low light, PRISM alpha^(B) (39.36 dB) and FisherRF (39.31 dB) differ by +0.05 dB, so the two are level there and no low-light advantage over FisherRF is claimed (Abstract, Sections 1, 4.1, 5). In addition, the sentence in the submitted Table 1 caption stating that random and coverage planners trail every row is removed because the curves do not support it on every testbed.

Figure status across the paper:

  • Schematic: Figure 1 (concept sketch; the caption now states that the bar values in the right panel are illustrative), Figure 2 (pipeline overview of one round; the operational steps are written out at the end of Section 3.5 and in Algorithm 1, Appendix B.5), and Figure 5 (comparison of the two nuisance-estimation paths, Appendix B.4).
  • Measured experimental outputs: Figures 3, 4, and 6-10. Figure 3 shows the held-out frame and focus region with the largest margin over the best baseline on each testbed, as stated in its caption, while full-dataset means are in Table 1. Figure 4 plots measured speed profiles on the UE5 route, with its scalar control clarified in Section 4.3 and the caption as the unmarginalized joint information I_aug. Figures 6, 7 and 9 are the residual diagnostics and stored coalition tables, Figure 8 the closed-loop curves for 2DGS and reflective scenes, and Figure 10 shows a measured single-candidate example on toycar (arriving image, render, per-ray G/B/V marginal maps, residuals normalized by sigma_0^2 and by c_r, and nuisance evidence).

A numerical log of one round is not included in the PDF; the code, configuration files, and candidate pools in the supplementary material can regenerate the coalition tables and round records.

W3 / Q3. Agreed: in the submission, Table 6 (submitted Table 5) compared allocation rules with Psi = 0 (and its fixed-order row tested one permutation, G, B, V, now stated in Section 4.4 and the Table 6 caption), Section 4.2 enabled online Psi only under uniform weights where persistent shares sum to I_pers, and the symmetric four-way rule was not run as a closed-loop planner.

Section 4.5 (Table 2) evaluates the four-channel design in one closed loop with one-hot task weights and online Psi on three public RobustNeRF scenes with real distractors (android, crab2, yoda; 3 seeds each, 9 paired runs per cell). Every planner uses beta = 0, so PRISM and the symmetric four-way planner differ only in the allocation rule. Online Psi comes from the mixture estimator using captured images only, with evidence parameters tau = 0.4 and pi_T = 0.05 fixed once on the clean yoda stream (Appendix C). Metrics: Task G is median relative depth error at SfM points (dimensionless, shown x10^-3, lower is better). Tasks B and V are measured against a reference model's rendering: Task B is base-color PSNR (dB, higher is better) and Task V is PSNR (dB) on the 20% of pixels with the highest view-dependent magnitude in the reference model.

Planner (9 paired runs/cell; PRISM wins in brackets)Task G (x10^-3)Task B (dB)Task V (dB)
PRISM (nested Shapley + Psi)28.020.016.9
Single-block information + Psi30.2 (7/9)19.6 (8/9)16.5 (8/9)
Rescaled singletons + Psi29.8 (7/9)19.7 (7/9)16.6 (8/9)
Symmetric four-way allocation29.9 (8/9)19.6 (6/9)16.5 (8/9)
No residual normalization (sigma_0^2 ref.)29.5 (7/9)19.7 (7/9)16.6 (6/9)
Oracle (ground-truth candidate mask)27.6 (3/9)20.1 (3/9)17.0 (2/9)

PRISM achieves the best mean among deployable planners on all three tasks, while the non-deployable ground-truth candidate-mask oracle is better on all three. PRISM outperforms single-block + Psi and rescaled singletons + Psi in 7/9 or 8/9 runs per task, and symmetric four-way allocation in 8/9 runs on Tasks G and V. On Task B vs. symmetric four-way (20.0 vs. 19.6 dB) and Task V vs. no residual normalization (16.9 vs. 16.6 dB), PRISM wins 6/9 runs with small mean differences. Conclusions are restricted to these three scenes (Section 4.5, Section 5).

W2 / Q2. Operationally, G, B, and V are persistent parameter blocks with rendering Jacobians J_b and Gauss-Newton block posteriors Sigma_b, so capturing a view reduces their uncertainty. By contrast, T is not a parameter block: it is a diagonal per-ray variance Psi(v) predicted from a per-primitive nuisance-evidence state q_k (updated from residuals of already captured views) and added to observation noise as R_T = sigma_0^2 I + Psi(v) when candidates are scored (Section 3.1). Thus T changes only the noise whitening in F_v(A; R_T) and supplies the conditional term S_T(v) = I(xi; Y_v | theta) (Eq. 11, Appendix A.2).

In the reported runs Psi enters the selection score only; the posterior is accumulated with sigma_0^2, so it is not nuisance-aware, and captured images are added to the standard training loss unmodified. Section 5 states this as a limitation.

Section 3.5 now ends with a five-step "One round" walkthrough connecting inputs, scoring, arrival, and updates (Algorithm 1 in Appendix B.5):

  1. Refit the Gaussian model on all captured views with the standard loss; re-accumulate block posteriors Sigma_b = (sum_r J_{b,r}^T J_{b,r} / sigma_0^2 + lambda_b I)^(-1) using fixed sigma_0 = 0.03 (512 sampled rays/view on mini-3DGS; 256 Rademacher probes/view on 2DGS).
  2. Project the current per-primitive nuisance-evidence state q_k onto the 384 scored rays (top 24 primitives per ray) of each candidate view v via rendering weights a_{r,k}: pi_r = min(1, sum_k a_{r,k} q_k) and Psi_r = pi_r tau^2 (Psi = 0 in static runs).
  3. Under R_T = sigma_0^2 I + Psi(v), evaluate the seven coalition log-determinants F_v(A; R_T) exactly from one shared Gram matrix, compute persistent shares (S_G, S_B, S_V) via Eq. (5) and S_T via Eq. (11), and select v* maximizing U_alpha(v) = alpha^T S(v) - beta S_T(v) (Eq. 6, which now drops an unused cost term and includes -beta S_T).
  4. When image v* arrives and before fitting, compute persistent reference scale c_r = sigma_0^2 + (1/3) tr([Phi_G + Phi_B + Phi_V]_rr) (Eq. 8) and frame-median factor s_t = max(med_r e_r^2 / (3 c_r), 1), evaluate the two-hypothesis weight gamma_r (Eq. 16), aggregate gamma_r onto visible primitives, and update q_k by an exponential moving average.
  5. Add v* to the training set unmodified for the next round.

Hand-set parameters (Section 3.3, Appendix C): one-hot task vectors alpha^(G) = (1,0,0), alpha^(B) = (0,1,0), alpha^(V) = (0,0,1) set per task (Reproducibility Statement corrected accordingly), default beta = 0, sigma_0 = 0.03, mini-3DGS prior stds (0.05 scene extent for position, 0.5 log-scale, 2.0 opacity logit, 0.3 base color, 0.3 direction coefficients), and evidence constants tau = 0.4, pi_T = 0.05, rho_min = 0.05 fixed once on validation data.

W1 / Q1. We agree that Gaussian IG, Shapley values, task weighting, and residual normalization are existing tools individually, and that conservation and uniqueness characterize the allocation rule rather than proving better acquisition (Theorem 3 shows only that under nuisance precedence and the axioms, the nested rule is not arbitrary; Appendix A.5).

What the decomposition adds over simpler task-specific scoring with nuisance marginalization is avoiding two mis-allocations of shared pixel evidence on the same posterior:

  • Against single-block task scoring with nuisance marginalization: because G, B, and V jointly render the same pixels, standalone block gains F_v({b}; R_T) hold the other uncertain blocks known and double-count shared evidence, summing to 1.5x I_pers on 3DGS scenes and 1.09x on 2DGS (Section 4.4, Figure 7c). On static tasks (Psi = 0, Table 6), singleton scores under identical one-hot weights trail PRISM on all three testbeds (2DGS: 4.88 vs. 4.79 x10^-3; low light: 39.02 vs. 39.36 dB; reflective: 17.38 vs. 17.51 dB) and also trail unweighted total information (4.85, 39.06, 17.44; Table 1). With online Psi enabled, the "Single-block information + Psi" and "Rescaled singletons + Psi" rows in Table 2 directly instantiate simpler task-specific scoring with nuisance down-weighting; PRISM outperforms both in 7/9 or 8/9 paired runs across Tasks G, B, and V (Section 4.5).
  • Against symmetric nuisance attribution: treating T as a fourth symmetric player evaluates persistent marginals with unobserved disturbance held known, inflating persistent credit by [Delta_v/4, 3 Delta_v/4] (Theorem 2). Table 2 isolates this rule at beta = 0, where PRISM wins 8/9, 6/9, and 8/9 runs.

Summary of changes: Figure 8 low-light panel removed and Table 1 caption corrected; joint closed-loop experiment added (Section 4.5, Table 2); five-step round walkthrough and Psi/posterior scope added (Sections 3.5, 5); numbering map: submitted Tables 2-5 -> Tables 3-6; new Tables 2, 7; new Section 4.5.

Thank you to all five reviewers for the time and care you put into reading and auditing our submission. Each review led to concrete improvements in the revision, We have answered each reviewer's points in detail in a separate response under their review, and we have uploaded the revised PDF.

Across the reviews, reviewers noted the clear motivation (jPkR, A8YP) and the theoretically grounded, conserving attribution (jPkR, TaW7, bqBa), with an asymmetric nuisance treatment that is well motivated (TaW7), analyzed by Theorems 2 and 3 (bqBa), and paired with a practical integration of nuisance evidence (bqBa) and residual-based transient segmentation using persistent uncertainty (c1bh). Reviewers also noted the flexible task adaptation (jPkR), thorough diagnostic experiments (TaW7), and better performance across different tasks from task re-weighting (c1bh).

The revision builds on these foundations with targeted experiments and presentation updates within the 10-page main text:

  • Joint closed-loop evaluation (Section 4.5, Table 2). We evaluate one-hot task weights with online Psi on three RobustNeRF scenes with real distractors (android, crab2, yoda), running 9 paired runs per cell with beta = 0 for every planner so the comparison with the symmetric planner isolates the allocation rule. PRISM achieves the best mean among deployable planners on all three tasks. It outperforms single-block + Psi and rescaled singletons + Psi in 7 or 8 of 9 paired runs per task, and beats symmetric four-way allocation in 8, 6, and 8 of 9 runs.
  • Single-block and task-utility comparisons. PRISM is ahead of single-block scoring on all three static tasks (Table 6: 4.79 vs. 4.88 depth error, x10^-3; 39.36 vs. 39.02 dB; 17.51 vs. 17.38 dB) as well as all three tasks of the joint test. In Table 1, each task's utility reaches the best mean on its matched metric, with the largest margins inside the task's focus region: +1.1 dB over total information on glossy segments, +0.6 dB on the darkest fifth of low-light frames, and 2.94 mm vs. 3.39 mm Chamfer distance on DTU scan 24.
  • Clarity and corrections. The revised PDF adds descriptive Section 3 titles, defines the four technical terms at initial use, ends Section 3.5 with a five-step "One round" walkthrough, positions six related methods in Section 2 and Table 7, and corrects Figure 8 and the captions.

Thank you all again for your time and effort. We wish you all the best.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.