acceptodds
Under review as a conference paper at ICLR 2027

EviSR-GS:Evidence-Guided Supervision and Refinement for Sparse-View Gaussian Splatting

12 views

Abstract

Sparse-view Gaussian splatting faces spatially uneven constraints from limited observations. Stronger pixelwise fitting can reinforce ambiguous reconstructions, while regions that could benefit from additional capacity may lack reliable cues for Gaussian placement. We propose EviSR-GS, an evidence-guided framework that combines spatially adaptive supervision with provisional Gaussian refinement, separating capacity allocation from subsequent activity assessment. For supervision, local depth residuals and image–depth gradient mismatch modulate the relative weights of pixelwise RGB reconstruction and patch-level perceptual losses. For refinement, low rendered coverage or disagreement with a fixed depth prior identifies candidate regions. Cross-view depth statistics and current coverage prioritize depth-based Gaussian proposals under a per-event insertion budget. Newly inserted Gaussians are jointly optimized with the existing representation and protected from ordinary pruning during a maturation period. Afterwards, additions with both low opacity and low accumulated rendering visibility undergo soft opacity attenuation. Experiments on LLFF, DTU, and Mip-NeRF 360 demonstrate competitive novel-view synthesis performance. On three-view LLFF, the full framework improves PSNR by 0.69 dB and reduces LPIPS by 0.021 relative to a baseline with both branches disabled. Ablations further support the contributions of both branches and the benefit of spatially varying loss weights over fixed scene-level weights. Code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Sparse-view Reconstruction, Novel View Synthesis, Spatially Adaptive Supervision, Provisional Gaussian Refinement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_HvApOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets sparse-view 3D Gaussian Splatting, where pixel fitting and densification can reinforce ambiguous scene explanations. EviSR-GS computes an input inconsistency map from local depth residuals and RGB-depth gradient mismatch, then uses it to balance pixel reconstruction and patch-level perceptual losses. A second branch proposes candidate Gaussians at low-coverage or depth-inconsistent locations, ranks them using cross-view depth statistics and rendered coverage, and inserts a limited number per event. New Gaussian lineages are protected from ordinary pruning during a maturation period. After maturation, additions with both low opacity and low accumulated visibility are softly attenuated. The method is evaluated on LLFF, DTU, and Mip-NeRF 360, with comparisons to sparse-view NeRF and Gaussian methods and additional input-view settings on LLFF.

Critical strengths
  1. The method treats local appearance supervision and Gaussian capacity allocation as related but distinct decisions, avoiding the assumption that every large image residual should receive stronger pixel fitting.
  2. The maturation period preserves newly inserted lineages while they adapt jointly with the existing scene representation. This is a clear response to the instability that can arise when new primitives are immediately judged by opacity or visibility.
  3. The refinement branch operates within the existing 3DGS pipeline and carries insertion, lineage, and activity state through ordinary densification and pruning events.
Critical weaknesses
  1. The supervision weight is a fixed combination of local depth variation and RGB-depth gradient mismatch, rather than a calibrated measure of depth correctness. Valid depth discontinuities, texture edges, or errors in the fixed prior can therefore reduce pixelwise constraints even when the measurements are reliable.
  2. Candidate construction requires finite positive rendered depth before applying the low-coverage trigger, and the paper explicitly says this excludes completely uncovered pixels. This leaves zero-coverage locations outside refinement even though low coverage motivates the trigger.
  3. Cross-view candidate statistics use nearest-neighbor depth comparisons without an additional occlusion mask. A candidate behind an occluder can appear unsupported, while a cross-view-consistent error in the fixed depth prior can appear well-supported even though agreement alone does not establish correctness.
  4. Post-maturation attenuation uses only current opacity and accumulated rendering visibility. These quantities describe contribution and exposure during training, but do not establish whether an addition represents a valid surface. Repeated attenuation can consequently reduce a useful but rarely visible structure.
Questions for authors
  1. How do the authors distinguish unreliable depth from valid depth discontinuities or texture edges that can produce similar signals in the supervision weight?
  2. If the fixed depth prior contains a systematic error that is consistent across views, can the cross-view support score distinguish it from a valid surface?
5 comments
AuthorsOpenReview ↗

Response to Reviewer HvAp: supervision, prior errors, and refinement

Thank you for distinguishing appearance supervision, capacity allocation, and the initial optimization period. The revision makes the role of each cue explicit and adds direct controls. All references below use the revised numbering.

W1/Q1 — Valid edges versus depth errors. U does not fully distinguish these causes; it measures inconsistency in the fixed RGB/depth inputs. We use it to balance exact pixel matching and local patch appearance, both against the same observed RGB. This preserves appearance supervision when pixel weights are reduced. Appendix A.1 bounds texture-induced downweighting: with zero local depth-residual energy, at least 74.7% of pixel weight remains even under maximal gradient mismatch. Valid depth discontinuities can still yield high U. The justification is the measured supervision benefit: spatial weighting improves PSNR/LPIPS by +0.14 dB/-0.005 with both losses retained (Table 2), and complete U removal increases three-seed mean LPIPS from 0.1622 to 0.1757 and reduces SSIM from 0.7602 to 0.7544 (Table 3(a)).

W3/Q2 — Systematic prior errors and occlusion. A wrong surface consistently represented in the fixed priors can receive high support. Agreement alone cannot distinguish that error from a correct surface. Our comparison uses nearest-neighbor sampling without an additional occlusion mask, so occlusion and boundary sampling can also increase discrepancy and demand. Table 12 quantifies selection sensitivity to source- and comparison-prior perturbations. Demand therefore specifies insertion priority, and the selected Gaussians remain subject to joint RGB optimization; support is not a geometric correctness test. Table 3(b) establishes the rendering benefit of the tested ranking and gate under the original priors.

W2 — Zero rendered depth. Both proposal triggers share the requirement for finite positive rendered depth. Consequently, zero-rendered-depth pixels are excluded even when the prior is valid and opacity is low. Section 3.3 states this immediately after Eq. (7). The implemented refinement addresses eligible low-coverage regions; it does not fill every completely uncovered location.

W4 — Rarely observed structures. Visibility counts rendering events, not distinct cameras. Ten visible renders from even one training view exclude an instance from attenuation. Low visibility alone is insufficient: the lineage must be mature and opacity below 0.03, subject to the active protection gates. This avoids penalizing a structure simply because few cameras observe it. A useful instance with both low opacity and fewer than ten visible renders can nevertheless be attenuated repeatedly, while a high-activity geometric error can escape attenuation. Appendix A.2 specifies these rules, including inherited lineage age and reset child visibility counts. Table 3(a) separately confirms the benefit of protecting immature lineages, improving the three-seed mean PSNR by 0.1411 dB with the assessment schedule fixed.

Reviewer_HvApOpenReview ↗

further questions

Thanks for the response. I still have two questions about the refinement part:

  1. In Table 15(a), the version without refinement actually looks better on SSIM/LPIPS. So what is refinement really adding?

  2. Most inserted points have no cross-view support, and attenuation rarely fires. So if a new Gaussian is placed at the wrong depth, what removes it?

AuthorsOpenReview ↗

Response to Reviewer HvAp: refinement contributions and handling of misplaced Gaussians

Thank you for the follow-up questions.

  1. Contribution of refinement. Refinement adds trainable capacity in regions with low coverage or depth disagreement. With the main spatial supervision retained, enabling refinement improves mean PSNR by 0.1780 dB, SSIM by 0.0019, and LPIPS by 0.0060 (Table 9(b)). This directly demonstrates its contribution alongside the supervision branch. Table 15(a) examines a separate configuration using full-image perceptual supervision without maturation protection. There, refinement improves PSNR by 0.0794 dB, with a small SSIM/LPIPS trade-off. The two comparisons establish an improvement across all three metrics under the main configuration and a more limited benefit under the supplemental configuration.
  2. Handling misplaced Gaussians. Newly inserted Gaussians undergo joint optimization of position and appearance. After maturation protection expires, ordinary pruning can remove eligible low-opacity additions, subject to the protection rules. Ordinary pruning remains active regardless of whether activity attenuation fires. Attenuation provides an additional reduction in the influence of mature additions with both low opacity and low visibility. An incorrectly placed Gaussian that remains active can survive, because the activity criterion measures rendering activity rather than depth accuracy. Our refinement mechanism therefore provides an optimization opportunity followed by activity-based management; its demonstrated benefit is improved held-out reconstruction in the complete method.
Reviewer_4n4oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

EviSR-GS addresses sparse-view Gaussian reconstruction through spatially adaptive supervision and provisional Gaussian refinement. The supervision branch uses local depth residuals and RGB–depth gradient mismatch to adjust pixelwise reconstruction and patch-level perceptual losses. The refinement branch inserts depth-based Gaussian proposals according to cross-view depth statistics and rendered coverage under a per-event budget. Inserted Gaussians are protected during a maturation period, after which additions with low opacity and low accumulated visibility undergo opacity attenuation. Experiments on LLFF, DTU, and Mip-NeRF 360 report competitive reconstruction quality, with ablations examining the supervision and refinement components.

Critical strengths
  1. The paper addresses a relevant problem in sparse-view reconstruction: local regions differ in the strength of their available constraints, so uniform supervision and capacity allocation may be suboptimal. Adapting both supervision and refinement to local evidence is a reasonable direction.

  2. The experimental evaluation includes useful controls beyond aggregate benchmark results. The uniform-weight comparison examines the contribution of spatial weighting while retaining both loss terms, and the candidate-selection experiments control insertion counts and maturation periods. These comparisons help distinguish the effects of individual design choices.

Critical weaknesses
  1. The methodological novelty is limited. While the ablations demonstrate empirical benefits, the inconsistency-based weighting and provisional refinement appear to be incremental extensions of existing depth-guided supervision and Gaussian-management strategies. Their substantive distinction from prior approaches remains unclear.

  2. The rationale connecting the evidence indicators to allocation and supervision decisions is insufficiently validated. RGB–depth mismatch may reflect valid texture, and weak cross-view support may reflect either limited observations or erroneous depth. Activity-based assessment does not directly resolve these ambiguities. The experiments establish overall improvements, but provide limited insight into whether the indicators guide the intended decisions.

Questions for authors
  1. How does the input-inconsistency indicator behave in texture-rich regions with reliable geometry, where RGB–depth gradient mismatch may be high despite accurate depth? What evidence supports reducing pixelwise supervision in such regions?

  2. Table 4 shows mixed results between demand-first and support-first selection. Under what conditions is demand-first preferable, and is there evidence that its additions improve poorly constrained geometry rather than mainly increasing appearance-fitting flexibility?

  3. Relative to NexusGS [1] and related sparse-view refinement methods, what is the main methodological distinction, and which controlled experiment most directly supports its benefit?

[1] Zheng, Yulong, et al. “NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting.” CVPR, 2025.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4n4o: Methodological Distinction and Supporting Controls

Thank you for the review. The revision adds controls that directly test supervision allocation, candidate ranking, and maturation protection. All references below use the revised numbering.

W1/Q3 — Methodological distinction. Building on NexusGS's depth priors, we make two coupled decisions: where pixel and patch appearance supervision is emphasized, and which depth-based additions receive a limited optimization opportunity. Inserted lineages are protected during initial optimization and assessed afterwards. The supporting controls directly test these decisions: spatial weights with both losses retained improve PSNR/LPIPS by +0.14 dB/-0.005 (Table 2); guided ranking with the same insertion cap reduces LPIPS by 7.4% versus random ranking after the gate (Table 3(b)); maturation protection with the assessment schedule fixed improves three-seed mean PSNR by 0.1411 dB (Table 3(a)). These isolate benefits of weighting, selection, and protection beyond adding a perceptual loss or unconstrained extra points.

Q1 — Texture and pixel supervision. U describes input inconsistency, not whether observed RGB is correct. When local depth-residual energy is zero, even maximal RGB-depth gradient mismatch retains at least 74.7% of pixel weight and cannot alone cross the regional threshold (Appendix A.1). This bounds the response to texture under the stated depth-statistic condition. A valid depth boundary can still receive high U. The design then changes the balance between exact pixel matching and patch appearance matching, with both losses supervised by the same RGB image. The controlled benefit is measured in reconstruction: Table 2 isolates spatial weighting, while the three-seed U-free control removes U-dependence from weighting, region construction, patch retention, and scheduling. Retaining U increases mean SSIM by 0.0058 and reduces mean LPIPS by 7.68% (Table 3(a)).

Q2 — When demand-first is preferable. In the matched three-seed replay, Demand-first gives higher mean PSNR under both assessment treatments. With Activity-soft, PSNR is 21.4936 versus 21.4510 dB, while Support-first gives slightly better SSIM/LPIPS, 0.7584/0.1606 versus 0.7582/0.1614 (Table 5). Demand-first prioritizes low coverage and weak or conflicting depth support; Support-first prioritizes placements corroborated by more views. The latter may help preserve local consistency, explaining a plausible source of the metric trade-off. The demonstrated preference for Demand-first concerns mean pixel accuracy in this setting, not every metric or scene.

These RGB results do not establish that the added capacity recovers more accurate surfaces. Weak support can reflect missing observations or erroneous priors, and high-activity misplaced Gaussians can escape attenuation. Section 3.3 and Appendix A.2 clarify these mechanisms; Table 12 characterizes prior-depth sensitivity. The contribution supported by the controls is improved allocation of supervision and optimization capacity for novel-view reconstruction.

Thank you for the response and additional controls. They strengthen the evidence that the proposed components are useful, but my concern about the significance of the contribution remains.

Spatial weighting improves PSNR by 0.14 dB over uniform weighting. More importantly, the direct comparison with support-first selection shows nearly identical performance: demand-first gains only 0.0426 dB in PSNR, with slightly worse SSIM and LPIPS. This comparison is particularly relevant to establishing the added value of the proposed candidate-selection rule. Although the random-ranking and maturation controls provide useful supporting evidence, they do not resolve this concern.

I therefore maintain my weak-rejection recommendation.

AuthorsOpenReview ↗

Response to Reviewer 4n4o: methodological contribution and scope of the ranking comparison

Thank you for acknowledging the stronger experimental evidence. We agree that Table 5 shows comparable performance between the two ranking strategies, with a small metric trade-off.

The methodological distinction we would like to emphasize is how uncertain evidence informs two different optimization decisions. For supervision, input inconsistency adjusts the balance between exact pixel matching and local patch appearance. For refinement, coverage and depth disagreement determine where additional capacity receives a limited optimization opportunity. Candidate placement is then optimized jointly with the existing representation and managed through maturation and activity assessment. This explicitly separates the need for additional capacity from confidence in its initial placement.

The supervision contribution has its own controlled evidence. Keeping both losses and their coefficients unchanged, spatial weighting improves SSIM from 0.756 to 0.766 and LPIPS from 0.162 to 0.157, alongside the 0.14 dB PSNR gain (Table 2). The three-seed U-removal control further shows a 7.68% LPIPS reduction from retaining U-guidance while both losses remain enabled (Table 3(a)).

These controls support a contribution beyond introducing a perceptual loss. We hope this clarifies how the supervision design and provisional capacity allocation contribute to the method, alongside the ranking trade-off shown in Table 5.

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

This paper proposes EviSR-GS, a sparse-view 3DGS method built on NexusGS. It adds two new components:

  1. Spatial supervision. A per-view indicator U is computed once from the fixed RGB image and the NexusGS depth prior. It combines a normalized 3x3 local depth variance with an RGB/depth gradient-magnitude mismatch. U de-prioritize the pixel MSE loss through Q = 1 - U and prioritize the patch-level VGG-LPIPS loss, which separates the image into covered and uncovered regions.

  2. Provisional refinement. Pixels with low rendered opacity or a large rendered-prior depth discrepancy are back-projected using the prior depth. Each candidate gets a demand score that favors fewer agreeing views, larger cross-view discrepancy, and lower coverage. Up to 1024 candidates are inserted per event. Inserted lineages are protected from pruning during a maturation period. Afterward, their opacity is multiplied by 0.1 if both opacity and a cumulative visibility count are low.

Experiments on LLFF (3 views) and DTU (3 views), and on MipNeRF360 (24 views), report the best LPIPS across all three benchmarks and the best PSNR/SSIM on LLFF and MipNeRF360.

Critical strengths
  1. Reasonable motivation. Two of the paper's premises are sensible and in line with current work on sparse-view GS reconstruction. Supervision should adapt to the reliability of each region. The newly added capacity should get a trial period before it is judged.

  2. Strong LLFF performance. The method reports strong LLFF 3-view numbers (21.68/0.766/0.157), which are the best performance. The proposed components are also plug-and-play modules for NexusGS, making it easy to validate their effectiveness.

  3. Detailed implementation and clean control. The implementation is stated in detail. The ablations are well organized, providing clear evidence of the contributions of spatial supervision, refinement, and weighing.

Critical weaknesses

W1. I do not find the core "evidence" logic convincing. The hand-crafted signals do not measure what the downstream decisions need, and worse, the two branches treat the same depth prior in contradictory ways.

(a) The refinement branch inserts Gaussians at the least trustworthy depths, and its later check cannot catch the resulting errors. Plugging the numbers into Eq. 8, a candidate whose prior agrees with both comparison views (nj=2n_j = 2) reaches at most ρj≤0.733<τρ=0.75\rho_j \le 0.733 \lt \tau_\rho = 0.75 and can never be inserted. A candidate with no agreeing view (nj=0n_j = 0) automatically passes with ρj≥0.8\rho_j \ge 0.8. New Gaussians thus end up at the least corroborated prior depths, which goes completely against the supervision branch that distrusts the same prior wherever it looks inconsistent. The cross-view check itself is weak evidence. It has no occlusion test, and NexusGS triangulates the priors from flow between the same training views, so agreement largely reduces to a forward–backward flow-consistency check. The post-maturation check does not fix this. A misplaced Gaussian that still fits the training pixels, exactly the geometry-appearance compensation that motivates the paper, has high opacity and visibility, and is simply kept. The mechanics are also more unstable than described. Attenuated opacity already falls below the pruning threshold, and if ν\nu is the standard visibility counter, matured insertions would rarely meet the visibility condition at all.

(b) The supervision "evidence" UU basically acts as a depth-edge detector rather than a reliability measure. The gradient mismatch term adds at most 0.25<τu=0.30.25 \lt \tau_u = 0.3 (Eq. 10), so crossing the threshold essentially requires large local depth variance, i.e., true depth discontinuities or a noisy prior. UU thus confuses clean occlusion boundaries, where the observed RGB is completely reliable, together with noisy priors. I do not see why loosening pixel fitting there would reduce geometric ambiguity, and the paper never checks whether UU tracks prior-depth error or novel-view error. Combined with (a), this means the conceptual story of evidence-guided allocation and assessment does not hold up, regardless of the benchmark scores.

W2. I am not convinced by the experimental evaluation either. The results do not support the claims about the individual mechanisms, and much of the gain is plausibly explained by adding a VGG-LPIPS training loss and comparing against a weakened baseline.

(a) The largest single effect comes from the patch LPIPS loss, measured against a weakened baseline. Dropping this loss costs more than removing any other component, and by far the most in LPIPS (Table 8b). Since the same VGG-LPIPS is used for training and evaluation, the "lowest LPIPS" claim partly reflects directly optimizing the evaluation metric. Meanwhile, Base replaces NexusGS's L1 + D-SSIM objective with MSE only and scores below the original NexusGS, so the +0.69+0.69 dB headline gain is measured against a weakened baseline. The obvious control, NexusGS plus a uniform patch LPIPS loss without refinement, is missing, and no UU-free variant is reported.

(b) The refinement mechanisms are not supported by the ablations. Most ablations appear to be single runs, while the seed standard deviation reaches 0.0360.036 dB (Table 4). Against this noise, several refinement components lower the PSNR when added. The demand threshold barely matters because the insertion budget is saturated, and the Q3 differences sit inside seed noise. All four Q3 variants even appear to score below the model without any refinement. The defining "provisional" ingredient, maturation protection, is never ablated, and there is no geometry-level check, such as the depth error of inserted and attenuated Gaussians on DTU. As a result, the current evidence does not separate demand-guided allocation, maturation protection, and activity-based attenuation from simply adding an LPIPS loss and a batch of depth-initialized Gaussians.

Questions for authors
  1. Please report the distribution of njn_j among inserted Gaussians at the default τρ=0.75\tau_\rho = 0.75, including how many have no valid comparison at all, and confirm that a candidate agreeing with both comparison views (nj=2n_j = 2) can never pass the threshold. If most insertions are unverifiable or contradicted points, the cross-view "evidence" framing does not hold. (W1a)

  2. How exactly is ν\nu computed (e.g., radii >0\gt 0), and what is its typical value for direct insertions at the end of maturation? Per assessment, how many instances are attenuated, how many of them are direct insertions vs. fresh clone/split children, and how many are pruned at the next density-control step? If attenuation primarily affects newly added children and is immediately followed by pruning, the "activity-based assessment of matured additions" story needs to be revised. (W1a)

  3. Does Q3 use the same supervision branch as Full? If so, why do all four Q3 variants score below the model without any refinement, and what accounts for the remaining gap to Full (seed, demand-dependent initial opacity, or scheduling)? (W2b)

  4. Full in Table 5 reports 49,501 direct insertions with a mean demand of 0.8208, while the τρ=0.75\tau_\rho = 0.75 row in Table 9 reports 50,760 and 0.8240, yet both have identical metrics (21.6820 / 0.7660 / 0.1570). Are these from the same run?

  5. Please add (i) NexusGS plus a uniform patch LPIPS loss without weighting or refinement, (ii) a UU-free variant with no UU-based masks, patch retention, or scheduling, (iii) NexusGS with its original L1 + D-SSIM objective reproduced in your codebase, and (iv) random candidate selection from the same triggered pool under the same budget. If (i) recovers most of Full's gain, or (iv) matches demand-guided selection, the evidence-guided components add little. (W2a)

  6. Please report Tm=0T_m = 0 (no maturation protection) vs. the default, and mean ±\pm std over at least three seeds for Tables 2 and 8. Without these, the contributions of maturation and of the individual refinement components cannot be distinguished from seed noise. (W2b)

  7. On DTU, please report the rendered-depth error of Base vs. Full, the depth error of inserted, retained, and attenuated Gaussians, and the correlation between UU and prior-depth error. This would directly test whether selection, attenuation, and UU track geometric reliability as claimed. (W1a, W1b)

4 comments
AuthorsOpenReview ↗

Response to Reviewer DWTv (1/3): Perceptual-Loss Controls, U-Free, and Maturation

Thank you for the detailed review. We added controls that separate supervision allocation, candidate selection, and maturation protection from the presence of a perceptual loss. All references use the revised numbering.

W2a/Q5(i, iii) — Original objective and plain LPIPS. Base is our common MSE ablation anchor; the 0.69 dB gain over Base is not presented as the gain over original NexusGS. Table 10 separately reports the earlier original-objective NexusGS reproduction (21.0221/0.7373/0.1781), NexusGS plus uniform LPIPS (21.0735/0.7406/0.1705), and main Full (21.6820/0.7660/0.1570).

The added control uses uniformly averaged full-image VGG-LPIPS, coefficient 0.25 every six iterations after iteration 500, with U-based supervision and refinement disabled. This configuration does not recover Full's reported performance. It tests full-image perceptual supervision; the weighting-only and U-free comparisons provide separate controls that retain the method's pixel and patch losses. Because LPIPS is also used during training, we report PSNR and SSIM alongside it and use the controlled differences to assess the additional design choices.

Q5(ii) — Complete U removal. U-free retains pixel and patch-perceptual losses, online refinement, and maturation protection. All pixel and retained-patch weights become one; regions use coverage alone; every grid patch intersecting its regional mask is retained; scheduling uses the fixed base schedule. No U-dependent mask or retention rule remains. Across three seeds, U-guidance improves mean PSNR/SSIM from 21.5112/0.7544 to 21.6885/0.7602 and reduces LPIPS from 0.1757 to 0.1622 (Table 3(a)). This supports U-guidance beyond simply including a perceptual loss.

Q5(iv) — Random selection. Table 3(b) samples uniformly without replacement after the demand gate, with the same 1,024-point insertion cap. Guided selection improves mean PSNR/SSIM/LPIPS from 21.6337/0.7593/0.1634 to 21.7235/0.7607/0.1513. The no-gate control retains score-based ranking and fills the cap at all 413 recorded events, yet retaining the gate lowers mean LPIPS by 8.0%. A saturated cap therefore does not make the eligibility rule irrelevant: it still changes which candidates can be selected.

W2b/Q6 — Maturation and training seeds. Table 3(a) directly switches immature-lineage pruning protection off/on while retaining the assessment schedule. Enabling protection improves the three-seed mean PSNR by 0.1411 dB, SSIM by 0.00152, and LPIPS from 0.16442 to 0.16220. The table reports mean and sample standard deviation across seeds 0, 1, and 2, each averaged over eight scenes. This establishes the reported mean benefits of protection and U-guidance across repeated training runs. The branch and weighting-only ablations remain seed-1 comparisons. Three-seed stability is not established here for every detailed component; the discussion of Table 9 describes the non-monotonic changes during progressive integration.

AuthorsOpenReview ↗

Response to Reviewer DWTv (2/3): Supervision Rationale, Candidate Support, and Geometry

W1 — Roles of the signals. Spatial supervision determines how observed RGB constrains fitting; refinement determines which depth-based candidates receive additional optimization capacity. U measures input inconsistency, demand specifies insertion priority, and opacity/visibility measures training activity. Sections 3.2-3.3 now state these roles explicitly.

W1b — Why reduce pixel weights? With sparse views, precise pixel fitting alone does not uniquely determine scene geometry. Our design retains the same observed RGB target while shifting relative emphasis from exact per-pixel matching toward local patch appearance in high-U regions. This does not treat texture or observed RGB as erroneous. When local depth-residual energy is zero, even maximal gradient mismatch retains at least 74.7% of the pixel weight and cannot alone cross the regional threshold of 0.3 (Appendix A.1). Valid depth boundaries can still produce high U. The supervision policy is tested by the weighting-only control (Table 2: +0.14 dB PSNR, -0.005 LPIPS) and the complete three-seed U-free control (Table 3(a)).

Q1 — Inserted support distribution. Your threshold calculation is correct: two-view-supported candidates have demand at most 0.7333 and fail the 0.75 gate; candidates without valid comparisons have demand at least 0.8 and pass, subject to subsequent ranking. Table 11 now reports all categories across 413 events: 24.9201% no valid comparison, 75.0765% valid comparisons without agreement, 0.0034% one agreeing view, and 0% two agreeing views. The rule therefore allocates provisional capacity predominantly to weakly supported locations. Joint optimization against RGB assesses their usefulness for reconstruction; the gate does not certify their depth.

Prior errors and Q7. Nearest-neighbor comparisons have no separate occlusion mask. Occlusion, boundary sampling, or wrong priors can increase demand, and a systematic error shared by the priors cannot be identified from agreement alone. Table 12 quantifies selection sensitivity: under gating, local -20% source-depth perturbation produces 49.62% overlap with the original selected-pixel sets, and boundary substitution produces 21.02%.

The requested independent DTU depth-error measurements and U/depth-error correlations are not available in this revision. The rendering controls establish improvements at held-out viewpoints, but do not separate correct-surface recovery from additional appearance-fitting flexibility. Section 3.3 explicitly states that an incorrectly positioned Gaussian can escape attenuation when its opacity or visibility reaches the threshold. Our supported claim is that the proposed supervision and allocation decisions improve novel-view reconstruction, rather than that these proxies identify every geometric error.

AuthorsOpenReview ↗

Response to Reviewer DWTv (3/3): Activity Assessment, Replay Protocol, and Result Provenance

Q2 — What is counted and attenuated. Appendix A.2 defines visibility as the number of main training renders marking an instance visible, including repeated observations from the same view. Direct insertions and clone/split children start at zero. Children inherit lineage birth time but not their parent's visibility count, so a fresh child of a mature lineage can be assessed. We now describe the targets as instances of mature tagged lineages, rather than implying that every assessed instance has itself completed the maturation period.

Attenuation requires opacity below 0.03 and fewer than ten visible renders, after maturity and protection checks. Multiplication by 0.1 puts opacity below 0.003, beneath the ordinary pruning threshold of 0.005. Ordinary pruning precedes assessment within the iteration; attenuation itself changes opacity without deleting points. Later optimization and pruning determine survival. The modified-protection diagnostic in Table 15(b) records 8/433 assessments with targets and 13 target selections, counting repeats. The requested typical-visibility, direct-insertion/child, and next-pruning statistics are not available for this response, so these counts do not establish which population dominates attenuation.

Q3 — Replay versus Full and no refinement. Yes, Q3 retains spatial supervision. Table 5 uses frozen candidate pools, matched actual insertion counts, common fixed initial opacity, and matched maturation schedules. Demand-first ranks stored demand alone. The online method instead constructs candidates from the evolving model, uses combined-score ranking, and initializes opacity from demand. The resulting absolute scores therefore compare different protocols; the gap has not been decomposed into individual causes.

Within the replay, Activity-soft improves all three means under both selectors, with PSNR gains of 0.0358 and 0.0603 dB. Their 0.0245 dB difference is reported descriptively in Table 14, without claiming a statistically established interaction. A separate matched online/no-refinement comparison in Table 15(a) improves PSNR by 0.0794 dB, with SSIM -0.0006 and LPIPS +0.0022. We report this metric trade-off rather than attributing the main Full result to refinement alone. Table 15 uses its supplementary supervision settings without age-based pruning protection; Table 3(a) separately isolates that protection.

Q4 — Identical displayed metrics with different allocation statistics. These are independent runs with different random seeds, as clarified in both captions. Table 6 reports 49,501 direct insertions and mean demand 0.8208; Table 13's default-threshold run reports 50,760 and 0.8240. Their scene-averaged reconstruction metrics coincide at the displayed precision. They are not allocation statistics from the same run.

Thank you for the detailed response and the new controls. A brief update on where I stand.

Resolved. The Q5(i)/(iii) control answers my W2a. With the original objective reproduced (21.0221/0.7373/0.1781) and NexusGS + uniform LPIPS (21.0735/0.7406/0.1705) against Full (21.6820/0.7660/0.1570), uniform perceptual supervision alone recovers only a small part of the gain. I withdraw the claim that the headline improvement is mostly a perceptual loss measured against a weakened baseline. The three-seed U-free control (21.5112/0.7544/0.1757 → 21.6885/0.7602/0.1622) and the three-seed maturation result (+0.1411 dB PSNR) are the right experiments and partly answer W2b and Q6. I also credit the authors for Table 11 and for stating the limits of their evidence plainly.

Remains. My soundness concern is narrowed but not removed, because the response confirms rather than rebuts the core of W1.

  1. The demand gate is inverted, now with numbers. The authors confirm my arithmetic and report that across 413 events, 24.92% of insertions have no valid comparison, 75.08% have comparisons without agreement, 0.0034% have one agreeing view, and 0% have two. They state that "the gate does not certify their depth". So provisional capacity is allocated exclusively to locations the prior fails to corroborate, while the supervision branch simultaneously distrusts that prior where it looks inconsistent. This is the contradiction I raised, confirmed empirically, and the "evidence-guided" framing does not survive it.
  2. No geometry-level evidence (Q7). The DTU depth-error measurements and the U/prior-error correlation "are not available in this revision", and the authors acknowledge the rendering controls "do not separate correct-surface recovery from additional appearance-fitting flexibility". Whether U or demand track geometric reliability — the paper's central premise — therefore remains untested. The W1b concession that gradient mismatch alone cannot cross τ_u = 0.3 (≥74.7% pixel weight retained) and that "valid depth boundaries can still produce high U" confirms that U conflates reliable occlusion boundaries with noisy priors.
  3. Activity-based assessment is close to inactive. The diagnostic reports targets in 8 of 433 assessments and 13 selections including repeats, and the per-population statistics "are not available for this response", so it is still unknown whether attenuation mainly hits fresh children that pruning would remove anyway. A mechanism that fires on ~2% of assessments is hard to credit as a contribution.
  4. Refinement's measured effect is marginal or negative. The matched online/no-refinement comparison gives +0.0794 dB PSNR with SSIM −0.0006 and LPIPS +0.0022, and the replay gains (0.0358 / 0.0603 dB) sit at roughly one seed standard deviation (0.036 dB). The authors state the Q3 gap "has not been decomposed into individual causes". The second of the paper's two contributions is thus not separated from noise.
  5. Result provenance (Q4) is not resolved by the explanation given. Table 6 and Table 13 are said to be independent runs with different seeds, yet they agree to four decimals on all three metrics (21.6820 / 0.7660 / 0.1570) while differing in insertions (49,501 vs 50,760) and mean demand (0.8208 vs 0.8240). With a reported seed standard deviation of 0.036 dB, two independent seeds coinciding at that precision is not plausible. Either the caption is in error or the seed variance is misstated; I would ask the authors to re-check. Relatedly, Full's own LPIPS appears as 0.1570, 0.1622 and 0.1513 across the main paper and the two rebuttal tables — a spread of 0.011, larger than several of the component effects being claimed.

I therefore keep my original ratings. The engineering is careful, and the authors responded honestly, including where the evidence does not support them. But the paper is framed as evidence-guided allocation and assessment, and after the rebuttal the gate provably selects uncorroborated points, the assessment mechanism almost never fires, and no geometric validation exists. The rendering gains are real; the explanation offered for them is not established. I keep soundness at 1 for this reason, and defer to the AC.

Reviewer_R5tTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EviSR-GS addresses sparse-view novel-view synthesis through spatially adaptive supervision and provisional Gaussian refinement. Using NexusGS depth priors, it adjusts pixelwise and patch-level perceptual supervision according to local RGB–depth inconsistency. Refinement allocates Gaussians to regions with low coverage or depth disagreement, protects inserted lineages during maturation, and subsequently attenuates additions with low opacity and visibility. Experiments cover LLFF, DTU, and Mip-NeRF 360. Ablations examine spatial weighting, candidate selection, and activity assessment.

Critical strengths
  1. A useful formulation of spatially uneven supervision and capacity allocation. The framework distinguishes regions needing additional capacity from regions providing reliable placement evidence. Spatial supervision and deferred assessment of inserted Gaussians offer a practical response to uneven constraints in sparse-view reconstruction.
  2. Positive internal evidence supported by meaningful controls. Both branches improve the common baseline. The uniform-weight control supports an additional benefit from spatial weighting: +0.14 dB PSNR and −0.005 LPIPS. The refinement comparison uses shared candidate pools, matched insertion counts and maturation periods, and three paired seeds. These controls and the held-out test viewpoint example strengthen the empirical case for the overall framework.
Critical weaknesses
  1. Comparative evaluation remains insufficiently controlled and incomplete. Baseline numbers use potentially different protocols, leaving small performance advantages difficult to interpret. Equivalence of LPIPS implementations and preprocessing is not established. Table 3 omits available six- and nine-view TWINGS results, which report LPIPS of 0.106/0.085 versus 0.112/0.086 achieved by the proposed method in this paper (see Table 3 in the original TWINGS paper). Moreover, it's weird that some results do not match with those from the corresponding original paper(s). For instance, in Table 3 of this submitted paper, CoR-GS achieves 24.34/0.831/0.122 for PSNR/SSIM/LPIPS for 6 views, while those numbers are 24.49/0.837/0.115 in Table 1 as in the original CoR-GS paper. And the same observation is found for 9 views.
  2. Evidence for the specific refinement rules is weaker than evidence for the overall framework. At the default threshold, candidates with two agreeing views are excluded because their demand cannot exceed 0.7333, whereas candidates without valid comparisons receive at least 0.8 and pass the gate. Robustness of this allocation policy to erroneous depth and occlusion is not demonstrated. Removing demand guidance costs only 0.0389 dB, and the selection–assessment interaction is 0.0245 dB without paired uncertainty estimates. Support-first achieves slightly better final SSIM and LPIPS. The opacity-only experiment does not isolate the visibility criterion’s added value. Furthermore, Section 3.3’s combined-score selection and Appendix B.3’s demand-only replay ranking require reconciliation.
Questions for authors
  1. Can you specify exact splits, resolutions, DTU masks, and LPIPS preprocessing, establish comparable evaluation against strong baselines, and include TWINGS’ six- and nine-view results?
  2. Is excluding candidates supported by both comparison views while admitting candidates without valid comparisons intentional? What are the selected proportions, and what evidence supports robustness to depth errors and occlusion?
  3. Does Table 4 use Section 3.3’s threshold and combined score or Appendix B.3’s demand-only ranking? Please clarify candidate-pool generation and initial opacity, and provide paired uncertainty for the interaction and multi-seed evidence for spatial weighting.
  4. Does opacity–visibility attenuation outperform opacity-only attenuation under identical conditions? How often does visibility affect decisions, and what training-time and peak-memory overheads do the branches introduce?
2 comments
AuthorsOpenReview ↗

Response to Reviewer R5tT (1/2): Corrected Comparisons and Demand-Gate Evidence

Thank you for the careful review. We corrected the baseline comparisons and added direct controls for demand gating, candidate ranking, U-guidance, and maturation protection. All references below use the revised manuscript's numbering.

W1/Q1 — Comparisons and evaluation. Table 4 now includes TWINGS at six and nine views: 24.97/0.844/0.106 and 26.23/0.876/0.085 (PSNR/SSIM/LPIPS). TWINGS has lower LPIPS in both settings. We replaced the secondary-source CoR-GS entries with the original-paper values, 24.49/0.837/0.115 and 26.06/0.874/0.089, and corrected the rankings and accompanying discussion.

LLFF uses eight scenes and the images_8 image set, with every eighth image held out. In our DTU evaluator, masks are reduced to one-quarter of their original dimensions and rendered/reference images are resized accordingly. PSNR uses foreground pixels; SSIM and LPIPS use the images after white-background compositing. Evaluation passes [0,1] PNG tensors directly to repository-local lpipsPyTorch VGG-LPIPS v0.1, retaining its internal normalization without an extra normalize=True flag or 2x-1 transform. External baseline numbers remain results reported by their source papers; our within-study controls share the evaluator. We distinguish these two comparison types in Appendix B.1.

W2/Q2 — Why the gate favors low support. This is intentional: demand assigns additional optimization capacity to locations with weak support, depth disagreement, or low coverage. It does not estimate depth confidence. With the default threshold 0.75, two-view-supported candidates cannot pass, while candidates without valid comparisons pass the gate and then compete in budgeted ranking (Section 3.3).

Table 11 reports the inserted composition: 24.9201% without valid comparisons, 75.0765% with comparisons but no agreement, 0.0034% with one agreeing view, and none with two. Table 3(b) tests the allocation decision directly. Keeping the gate reduces mean LPIPS by 8.0% versus ungated score-based ranking; guided ranking reduces LPIPS by 7.4% versus random selection after the gate. Mean PSNR and SSIM also improve under the same 1,024-point insertion cap.

Prior errors, occlusion, and boundary sampling can increase demand. Table 12 now measures this sensitivity at fixed checkpoints: local -20% source-depth perturbation gives 49.62% selected-set overlap, and boundary substitution gives 21.02%, under gating. These are selection-sensitivity measurements, not evidence of occlusion robustness. Together, Tables 3(b), 11, and 12 distinguish the measured rendering benefit, the candidates receiving capacity, and the dependence on source priors.

AuthorsOpenReview ↗

Response to Reviewer R5tT (2/2): Replay Protocol, Multi-Seed Results, and Visibility

Q3 — Online selection versus frozen replay. Table 5 (formerly Table 4) uses frozen candidates; its Demand-first arm ranks by stored demand alone. Section 3.3 describes the online selector, which applies the demand gate and ranks by the combined source-view and demand score. The replay pool is recorded after validity checks and source-view preselection, before strategy-specific filtering. The four configurations share the frozen pool, actual insertion counts, spatial supervision, fixed initial opacity, insertion/maturation schedules, and ordinary density-control settings. This shared opacity replaces the online method's demand-dependent initialization (Appendices B.2-B.3).

In the paired three-seed replay, Activity-soft improves all three reported means under both selectors. Demand-first gives higher mean PSNR under both treatments; Support-first gives slightly better final SSIM/LPIPS. Support-first favors placements corroborated by more views, which may help local structural consistency; this is a mechanism-based explanation of the observed trade-off. Table 14's 0.0245 dB difference is a descriptive mean interaction. We do not infer its statistical significance from per-configuration standard deviations; paired uncertainty for that interaction is not reported.

Multi-seed supervision evidence. Table 3(a) adds paired seeds 0, 1, and 2 for U-guidance and maturation protection. With protection enabled, U-guidance improves mean SSIM by 0.0058 and reduces LPIPS by 7.68% relative to U-free, while retaining both losses and refinement. This tests U-dependent weights, regions, patch retention, and scheduling together. The weighting-only control in Table 2 remains a seed-1 comparison, improving PSNR by 0.14 dB and LPIPS by 0.005. We distinguish these experiments rather than extending the three-seed claim to every component.

Q4 — Visibility condition. Table 15 compares the count condition with other settings matched within each panel. Under panel (a)'s contribution/texture protections, retaining it improves mean PSNR/SSIM/LPIPS by +0.1266 dB/+0.00109/-0.00047. Panel (b)'s modified protection settings favor opacity-only on all three means. Thus, the count condition helps in panel (a), with its benefit depending on the surrounding protection rules. These supplementary controls use their stated supervision settings and do not enable age-based pruning protection.

Visibility counts main-training-render events in which an instance is marked visible, including repeated views. Attenuation requires a mature lineage, opacity below 0.03, and fewer than ten visible renders, subject to the active protection gates. In panel (b), the joint rule selects targets in 8/433 assessments, totaling 13 selections including repeats. This measures joint-rule activity; the number of decisions changed specifically by visibility is not available.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.