acceptodds
Under review as a conference paper at ICLR 2027

DGS: Dual-Evidence Integration for Accurate Sparse-View Surface Reconstruction with Gaussian Splatting

Abstract

Accurate surface reconstruction with 3D Gaussian Splatting (3DGS) remains challenging under sparse views, where limited multi-view observations provide insufficient geometric constraints. Recent methods alleviate this issue by supervising reconstruction with aligned depth predictions from foundation models. However, these predictions typically lack explicit dense multi-view matching and can be suboptimal where reliable photometric correspondences are available. More importantly, their depth reliability is inherently range-dependent: aligned VGGT depth exhibits smoothly varying bias across surfaces, while local depth variations remain substantially more accurate and degrade gradually with spatial distance. Motivated by this, we propose DGS, a dual-evidence integration framework for sparse-view surface reconstruction. Rather than directly using predicted depth as supervision, DGS combines two complementary forms of geometric evidence: point evidence from multi-view photometric matching provides accurate absolute anchor depths at reliable correspondences, while edge evidence from local log-depth differences of VGGT captures dense local depth variations. A global integration couples the two, extending anchor depths through local edge relations while constraining accumulated drift. We further introduce prior-based aggregation and anchor depth refinement to improve the coverage and accuracy of point evidence under sparse views. Experiments on DTU and Tanks and Temples demonstrate state-of-the-art surface reconstruction quality.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Sparse-View Surface Reconstruction, 3DGS

Reviews

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

This paper studies sparse-view surface reconstruction with 3D Gaussian Splatting, using PGSR as its reconstruction backbone. Its motivation is that aligned feed-forward depth predictions can exhibit spatially varying absolute bias, while local depth differences remain considerably more accurate. DGS combines photometric depth anchors with local log-depth differences from a geometric prior to construct dense depth supervision.

For point evidence, the method evaluates 61 disparity-sampled depth hypotheses around the current rendered depth, aligns neighboring matching-cost profiles using prior depth ratios, aggregates them within a local window, and extracts reliable anchors with sub-index refinement. For edge evidence, it uses local log-depth differences after estimating and removing a global shift in the prior. A robust integration objective combines these constraints and is solved using IRLS and preconditioned conjugate gradients. The resulting depth supervises PGSR. Experiments on DTU and Tanks and Temples (TnT), together with component and supervision ablations, evaluate reconstruction quality and the proposed evidence construction.

Critical strengths

1. The use of geometric priors is motivated by a useful insight and supported across multiple priors. The paper distinguishes the reliability of absolute depth from that of local depth differences, using photometric anchors to establish absolute constraints and local differences to propagate structure. In Appendix Table 11, replacing VGGT with DA3, MoGe3, and π³ still improves CD over global alignment, from 0.965 / 1.150 / 0.947 to 0.766 / 0.820 / 0.757, respectively. This supports the broader relevance of the supervision strategy beyond one particular prior.

2. Reconstruction gains are consistent and supported by controlled comparisons. The method achieves a CD of 0.71 on DTU and an F-score of 0.3100 on TnT, outperforming the compared methods. Particularly compelling is Table 4: with the same backbone and optimization schedule, integrated depth achieves a CD of 0.707, compared with 0.907 for aligned VGGT depth, 1.059 for anchor-only supervision, and 0.806 for binary fusion. This comparison provides direct evidence for the benefit of constructing depth supervision through integration.

Critical weaknesses

1. The recovery range of anchors under inaccurate initialization remains insufficiently characterized. Depth hypotheses are centered on the current rendered depth, and Algorithm 1 / Table 8 accepts a matching minimum only when its index offset is at most τ=4. This helps reject unreliable matches, but may also reject candidates requiring a larger depth correction. Appendix A.3.4 shows that new anchors emerge as optimization progresses, so the method is not necessarily limited to the correction available in a single update. The remaining question is whether the matching–integration–reconstruction loop can recover from substantially inaccurate initial geometry, and how sensitive this behavior is to the acceptance threshold. This affects the conditions under which point evidence can supply reliable absolute geometric constraints.

PBA also uses prior depth ratios to align cost profiles, so point and edge evidence are not fully independent. Table 3 establishes the benefit of PBA as a whole, but does not separately establish the benefit of prior-guided alignment over ordinary local aggregation. Although this is compatible with avoiding direct supervision by absolute prior depth, it leaves uncertainty about whether errors in the prior's local structure can bias both evidence sources in the same direction.

2. The effective range of integration in anchor-poor regions and under spatially varying bias needs clearer qualification. Figure 6 and the conclusion report limited integration gains beyond approximately 32 pixels from an anchor in the evaluated setting. The >10-pixel group in Table 10 also shows no advantage over kernel regression, with 69.26% versus 69.63% of pixels below the stated error threshold; this small difference alone does not establish a significant degradation. These results indicate that local structure can propagate anchor information, but does not automatically resolve extensive regions without reliable matches. The detailed diagnostics are primarily on DTU. The aggregate TnT F-score improvement does not establish how much of each scene lacks anchor support or how accurate those regions are. The pixel-distance range observed on DTU should also not be assumed to transfer unchanged across image resolutions and scene settings.

In addition, Equation 9 approximates prior error using a shared scale and constant shift, whereas Figure 1 illustrates spatially varying bias. Global de-shifting can remove a constant component but does not eliminate spatial variation. The empirical local-difference analysis supports the approach, but the paper could more clearly explain why residual errors remain controlled over the anchor spacings encountered in practice. These concerns limit the established operating conditions rather than overturning the controlled gains in Table 4.

Questions for authors
  1. Clarification and targeted validation: how much initial error can the anchors correct? Under the existing DTU evaluation setup, could you group pixels by their initial rendered-depth error and report how depth error and anchor coverage evolve over successive updates? Please also clarify how τ=4 affects coverage, accuracy, and final CD. The concern is persistent rejection of correct matches under poor initialization, rather than the overall anchor accuracy already reported in Figure 6.

  2. Clarification: how prevalent are anchor-poor regions on TnT? For the four scenes in Table 2, please report anchor coverage and nearest-anchor distance distributions at specified update iterations, stating the image resolution and valid-pixel definition. These statistics do not require ground-truth depth. A scene-level comparison with a key alternative supervision strategy, such as aligned prior depth, using the same 10-view selection and F-score protocol would help connect coverage to reconstruction performance. Pixelwise depth-error diagnostics can remain on DTU; this request does not assume that TnT provides directly comparable dense depth ground truth or that DTU's pixel-distance and millimeter thresholds transfer to TnT.

  3. Clarification: how do local prior errors affect the two evidence sources? Does the PBA ablation in Table 3 include ordinary index-wise local aggregation? If not, please clarify the gain or failure cases associated specifically with prior-guided alignment. Please also explain how spatially varying residual bias after global de-shifting affects local differences and integration; residual-difference statistics on the existing data would help establish the validity of this approximation.

  4. Clarification of computational cost. Which time and memory entries in Table 5 were measured under common hardware and settings, and which were taken from other reports? Does the reported total include prior inference and depth updates? A brief breakdown of matching/PBA and IRLS/PCG within each depth update would help explain the reported 4.3 min / 6.0 GB, compared with VGGS's 1.8 min / 2.4 GB.

4 comments
AuthorsOpenReview ↗

Response to Reviewer cvjP (Part 1)

We sincerely thank the reviewer for the careful reading and the detailed, constructive feedback. We address each question below and will update the results in the final paper.

Q1: Clarification and targeted validation: how much initial error can the anchors correct?

We thank the reviewer for this insightful question. Following the suggestion, we conduct the analysis on DTU. For each foreground pixel, we measure its rendered-depth error at the first update (iteration 2k), when the depth supervision is introduced. We group the pixels by this initial error and track the same pixels over the subsequent updates (Tab. R1). We further vary the index-offset threshold, denoted as η in the paper, to examine its effect (Tab. R2). It is set to 2 on DTU and 4 on TnT, and Tab. 8 mistakenly lists only the TnT value, which we will correct.

  1. The rendered-depth error decreases in all groups. As shown in column (a) of Tab. R1, small and moderate errors are reduced to the sub-millimeter level (rows (1)–(2)), and initial errors above 4 mm are reduced to about 2 mm (rows (4)–(5)).
  2. The correction of large errors is initiated by the integration, after which anchors emerge. As shown in rows (3)–(5) of Tab. R1, pixels with large errors have few and inaccurate anchors at the first update, since the anchor search is centered at the erroneous rendered depth. The integration, constrained by the correct anchors in the surroundings and the local depth relations of the prior, first pulls the rendered depth closer to the true surface. Correct matches then fall within η and are accepted as anchors, and the anchor coverage and accuracy increase in all groups (columns (b) and (c)).
  3. Rejection by η is temporary, and η has limited influence. At the first update, some correct matches are rejected because they lie outside η. As the rendered depth moves closer to the true surface, these matches fall within η and become anchors, as discussed above. Relaxing η to 4 or 8 slightly increases the anchor coverage (column (a) of Tab. R2) at the cost of slightly lower accuracy (column (b)), and yields only a marginal change in CD (column (c)). This indicates that rejection by η is not a bottleneck of the reconstruction.

Tab. R1: Evolution of pixels grouped by their initial rendered-depth error on DTU (η = 2). Pixel ratio: percentage of foreground pixels in each group. Anchor coverage: percentage of pixels selected as anchors.

Initial error (mm)Pixel ratio (%)(a) Median rendered-depth error (mm, 2k → 6k)(b) Anchor coverage (%, 2k → 6k)(c) Anchors with error < 0.5 mm (%, 2k → 6k)
(1) [0.5, 1)24.50.71 → 0.3960.4 → 64.975.7 → 75.9
(2) [1, 2)17.91.35 → 0.6430.8 → 45.238.0 → 56.8
(3) [2, 4)8.72.61 → 1.128.0 → 22.70.0 → 45.6
(4) [4, 8)3.55.28 → 1.984.0 → 11.30.0 → 35.0
(5) ≥8\geq 82.010.69 → 2.192.4 → 7.50.0 → 34.2

Tab. R2: Effect of the index-offset threshold η on DTU.

η(a) Anchor coverage (%, 2k → 6k)(b) Anchors with error < 0.5 mm(c) CD ↓
2 (default)57.8 → 63.577.10.707
461.1 → 65.275.90.706
862.1 → 65.875.20.710
AuthorsOpenReview ↗

Response to Reviewer cvjP (Part 2)

Q2: Clarification: how prevalent are anchor-poor regions on TnT?

We thank the reviewer for this constructive comment. Following the suggestion, we report the anchor statistics on the four TnT scenes in Tab. R3, together with a scene-level comparison against supervision with the aligned prior depth under the same 10-view selection and F-score protocol. All statistics are computed at the training resolution of 980 × 520. Since the background is not evaluated on TnT and contains few anchors, we compute the statistics only on foreground pixels. The foreground mask is obtained by projecting the ground-truth point cloud into each training view with z-buffering, where each point is dilated to a 3 × 3 patch to fill the gaps between points. The mask is used only for this analysis.

  1. The anchor support varies across scenes. As shown in Tab. R3, Ignatius and Truck are well covered, with most pixels close to an anchor. Barn and Caterpillar contain more anchor-poor regions, where the 90th-percentile distance to the nearest anchor reaches about 20 px. In all scenes, the anchor coverage increases as the optimization proceeds.
  2. DGS improves the reconstruction in all scenes, and the gain tends to be larger with denser anchor support. In the scenes with better anchor coverage (Ignatius and Truck), DGS brings large improvements over the aligned prior depth. Even in the scenes with more anchor-poor regions (Barn and Caterpillar), DGS still outperforms using the aligned prior depth as supervision.

We agree with the reviewer that the pixel distances observed on DTU should not be assumed to transfer unchanged across resolutions and scenes, and that scenes with few reliable matches remain challenging (e.g., Caterpillar). Our analysis on DTU is intended to characterize the behavior of the integration in that setting, rather than to provide universal thresholds. We will clarify these conditions in the revision.

Tab. R3: Anchor statistics and F-score (↑) on Tanks and Temples with 10 input views.

SceneAnchor coverage (%, 2k → 6k)Mean distance to nearest anchor (px)P90 distance (px)< 5 px (%)< 10 px (%)< 15 px (%)F-score, aligned priorF-score, DGS
Barn12.5 → 25.98.222.153.970.480.30.15180.2371 (+0.0853)
Caterpillar9.9 → 19.38.119.046.867.881.30.09180.1331 (+0.0413)
Ignatius56.2 → 62.42.89.082.590.994.80.22540.3530 (+0.1276)
Truck23.4 → 33.94.511.764.384.193.70.32280.5168 (+0.1940)
Mean––––––0.19800.3100 (+0.1120)
AuthorsOpenReview ↗

Response to Reviewer cvjP (Part 3)

Q3: Clarification: how do local prior errors affect the two evidence sources?

We thank the reviewer for this question. We address it in two parts.

Effect of residual bias after de-shifting. We agree that the prior error varies spatially, and Eq. 9 is a simplification. In DGS, the de-shifting only estimates and subtracts the global shift component, since this component causes systematic errors in the local depth differences that share the same direction and accumulate with distance. Tab. R4 verifies the effect of the global shift. Compared with scale alignment, scale-shift alignment yields the same error at 1–2 px, and the gap between them increases with distance, showing that the de-shifting alleviates the accumulated error caused by the global shift. The remaining spatially varying component still affects the local depth differences, but its error is small over short distances (Tab. R4). During integration, such errors accumulate only between anchors and are pulled back at the anchors, which suppresses further accumulation. On DTU, 90% of pixels lie within 8 px of an anchor (Fig. 6c), where the residual error of the local depth differences remains below about 0.3 mm (Tab. R4).

Tab. R4: Local depth-difference error (mm, ↓) of the VGGT depth versus pixel distance on DTU. Scale alignment aligns the prior with a global scale only, without de-shifting. Scale-shift alignment additionally removes the global shift.

Pixel distance1 px2 px5 px10 px20 px50 px100 px
Scale alignment0.080.130.240.370.530.891.25
Scale-shift alignment (Fig. 1)0.080.130.220.330.480.750.93

Effect of local prior errors on the anchors. PBA uses the local depth ratios of the prior only within a 5 × 5 window, i.e., within 2 px, to align the cost curves of neighboring pixels. Over such a short distance, the relative depth of the prior is reliable, with an error of about 0.13 mm at 2 px (Tab. R4). Since the anchors are finally selected from the cost curves, the influence of local prior errors is further suppressed. The ablation in Tab. 3 does not include index-wise aggregation. We therefore compare PBA with ordinary index-wise aggregation, which averages the costs of neighboring pixels at the same hypothesis index without alignment (Tab. R5). Compared with per-pixel matching, index-wise aggregation already improves the accuracy and coverage of the anchors, and PBA brings further gains in all three metrics. This indicates that, within such a short range, the gain brought by the prior outweighs the influence of its local errors.

Tab. R5: Effect of cost aggregation on DTU.

AggregationAnchor coverage (%)Anchors with error < 2 mm (%)CD ↓
None (per-pixel)49.689.10.834
Index-wise aggregation55.091.60.748
PBA57.893.60.707

Q4: Computational cost

In each depth update, generating the supervision depth takes about 2.25 s per view. Anchor construction takes 1 s, including 0.6 s for building the cost volume and 0.4 s for cost aggregation. Integration takes 1.25 s, most of which is spent on the IRLS iterations with the conjugate-gradient solver (1.21 s). The additional memory mainly comes from building and aggregating the cost volume (about 3.5 GB), while the integration requires about 0.4 GB.

All 3DGS-based methods in Tab. 5 are reproduced and measured by us on a single RTX 3090. The results of NeuSurf and SparseCraft are taken from VGGS, where they were also measured on an RTX 3090. For all methods we measured, the reported time includes the depth updates and excludes the inference of geometric priors. The longer time of DGS compared with VGGS mainly comes from the number of iterations, as DGS is optimized for 6k iterations while VGGS converges within 3k iterations.

Thank you for the thorough responses and additional experiments. The controlled comparison between propagation and integration, using both anchor sources, clarifies the distinction from DP-GS and the complementary contributions of anchor construction and integration. The experiments across view counts, the dense-view comparison with a shared 2DGS backbone, and the anchor-removal studies also address my concerns about sparsity and anchor availability. I appreciate the additional ScanNet++ evaluation and the explicit discussion of failure cases involving textureless or reflective regions and noisy priors. These results provide a clearer picture of the method’s applicability and limitations. While I still view the contribution primarily as a well-motivated integration of established techniques, the additional evidence strengthens the case for the proposed design. I encourage the authors to incorporate these comparisons and limitations into the revised paper. Overall, my concerns have been substantially addressed, and I will maintain my weak acceptance rating.

Reviewer_b1EcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DGS, a sparse-view surface reconstruction framework that improves surface reconstruction quality by introducing a dual-evidence integration paradigm. The method is motivated by the biases of the predicted VGGT depths, while their local depth differences are more reliable. DGS introduces PBA and ADR to improve anchor coverage and accuracy, while an anchor-based de-shifting step reduces bias in the relative depth constraints. Experiments on DTU and Tanks and Temples report improvements over competing methods.

Critical strengths
  1. Clear Motivation and Analysis.

This paper presents a clear insight for combining photometric matching and the VGGT depth prior, addressing a strong analysis for the usage of the directly aligned VGGT depth.

  1. Comprehensive Experimental Evaluation.

The experiments are comprehensive, achieving consistent improvements on the evaluated benchmarks.

Critical weaknesses
  1. Unclear Attribution of Novelty.

The main source of novelty is unclear: the motivating insight emphasizes the integration framework, while the methodological innovations appear to focus on anchor construction through PBA and ADR. Although these components are ablated, comparisons with alternative anchor selection methods under the same integration framework would help clarify where the improvements come from. A convincing clarification of this issue would make me consider raising my score.

  1. Presentation Clarity.

The overall presentation could be clearer and more concise. The paper would benefit from a more focused explanation of its key ideas and contributions.

Questions for authors
  1. Source of Novelty.

Could the authors clarify whether the primary novelty lies in the integration framework or the anchor construction strategy, and provide evidence distinguishing their contributions?

  1. Sensitivity to Initialization.

How sensitive is DGS to initialization quality? Since anchor selection rejects matches far from the current rendered depth, can the method recover from substantial initial geometry errors, or does it rely on an accurate MASt3R-SfM initialization?

3 comments
AuthorsOpenReview ↗

Response to Reviewer b1Ec (Part 1)

We sincerely thank the reviewer for the detailed and constructive feedback. We address each question below and will update the results in the final paper.

Q1: Source of novelty

We thank the reviewer for raising this important question. We view the integration and the anchor construction as two complementary parts of a single design, which together realize a more effective use of geometric priors.

To distinguish their contributions, we conduct a controlled experiment on DTU. We compare two anchor selection schemes: (1) the DP-GS scheme, which selects anchors by thresholding a confidence score computed from NCC and reprojection error, and (2) our anchor construction. We further compare three densification schemes that combine the anchors with the VGGT depth prior to generate the supervision depth for 3DGS optimization: (a) global alignment, (b) DP-GS high-confidence propagation, and (c) our integration. All experiments use the same framework, initialization, and training losses. The NCC loss and multi-view consistency loss used in DP-GS are not included. The results are shown in Tab. R1.

  1. Contribution of the anchors. Following the suggestion, we compare different anchor selection methods under the same integration framework ((1c) vs. (2c)). Replacing our anchors with DP-GS anchors increases CD from 0.707 to 0.906, showing that the anchor construction is essential for the integration to take effect. Our anchors also outperform DP-GS anchors under DP-GS propagation ((2b) vs. (1b)) and global alignment ((2a) vs. (1a)), demonstrating that the advantage of our anchor construction holds across different densification schemes.
  2. Contribution of the integration. With the same anchors, our integration outperforms both the global alignment used in VGGS and the propagation used in DP-GS ((2c) vs. (2a) and (2b)). To further examine the shape of the supervision depth, we convert it into normals and compare them with the ground-truth normals (Tab. R2). Integration improves the normal accuracy over the VGGT depth, showing that it uses the accurate anchors to correct the erroneous shape of the prior. Propagation instead increases the mean normal error, as repeated averaging over-smooths the details of the prior.

The largest improvement is achieved when both are combined (2c), as the integration exploits dense and accurate anchors that global alignment cannot fully use, while such anchors in turn enable the integration to take better effect. We will clarify the respective contributions of the two parts in the revision.

Tab. R1: Chamfer Distance (↓) on DTU with different anchors (rows) and densification schemes (columns).

Anchors(a) Global alignment(b) DP-GS propagation(c) Our integration
(1) DP-GS anchors0.9191.2550.906
(2) Our anchors0.9070.9880.707

Tab. R2: Normal accuracy of the supervision depth on DTU.

Supervision depthNormal error (°) ↓Normal < 5° (%) ↑Normal < 15° (%) ↑
VGGT depth14.2722.669.3
Our anchors + DP-GS propagation15.6522.867.4
Our anchors + our integration13.5826.771.9

Q2: Presentation clarity

We thank the reviewer for this valuable comment. We believe the main source of confusion is that the paper does not explicitly connect the motivation to the anchor construction. Our analysis shows that the relative depth of the prior is reliable only over a short range. The integration therefore uses the prior only for local depth relations, and its accuracy depends on anchors that are dense and accurate enough to keep the prior within this range. PBA and ADR are designed precisely for this purpose. In the revision, we will make this connection explicit in the introduction and at the beginning of the method section, and present the integration and the anchor construction as two complementary parts of a single design rather than separate contributions.

AuthorsOpenReview ↗

Response to Reviewer b1Ec (Part 2)

Q3: Sensitivity to initialization

To evaluate the sensitivity of DGS to initialization, we initialize both DGS and VGGS with the COLMAP point clouds provided by VGGS. As shown in rows (a) and (b) of Tab. R3, COLMAP initialization, which is sparser and less accurate than the MASt3R-SfM point clouds, degrades the performance of both methods. However, DGS with the poorer COLMAP initialization still clearly outperforms VGGS with the MASt3R-SfM initialization.

We further perturb the COLMAP points with Gaussian noise of increasing levels (rows (b)–(e)). Noise with a standard deviation of 1 mm has a negligible effect on both methods. As the noise level increases further, both the accuracy and completeness of the two methods degrade. Notably, DGS is much less sensitive to noise. With 5 mm noise, its mean CD increases by only 0.081 mm (10%) relative to the COLMAP initialization, compared with 0.329 mm (31%) for VGGS. Moreover, even with 5 mm noise, DGS still outperforms VGGS with the MASt3R-SfM initialization, demonstrating its robustness to initialization noise. We will add this experiment to the appendix.

The above experiments evaluate the overall robustness to initialization. To directly examine whether DGS can recover from substantial initial geometry errors despite rejecting matches far from the rendered depth, we group pixels on DTU by their rendered-depth error at the first depth update (iteration 2k) and track them over the subsequent updates (Tab. R4). Pixels with large initial errors have few anchors at the first update, since the hypothesis search is centered at the erroneous rendered depth. They are nevertheless constrained by the integration, which extends the correct anchors in the surroundings through the local depth relations of the prior and pulls the rendered depth toward the surface. As the geometry improves, correct matches enter the search range and anchors emerge in these regions. As a result, initial errors above 4 mm are reduced to about 2 mm by iteration 6k.

Tab. R3: Chamfer Distance (↓) on DTU with different initializations.

InitializationVGGSOurs
Acc. ↓Comp. ↓Mean ↓Acc. ↓Comp. ↓Mean ↓
(a). MASt3R-SfM (default)0.6641.1620.9130.5040.9100.707
(b). COLMAP0.7911.3091.0500.5881.0260.807
(c). (b) + noise (σ = 1 mm)0.818 (+0.027)1.271 (−0.038)1.044 (−0.006)0.605 (+0.017)1.010 (−0.016)0.807 (+0.000)
(d). (b) + noise (σ = 3 mm)0.924 (+0.133)1.402 (+0.093)1.163 (+0.113)0.655 (+0.067)1.047 (+0.021)0.851 (+0.044)
(e). (b) + noise (σ = 5 mm)1.137 (+0.346)1.621 (+0.312)1.379 (+0.329)0.672 (+0.084)1.105 (+0.079)0.888 (+0.081)

Tab. R4: Evolution of pixels grouped by their initial rendered-depth error on DTU.

Initial error (mm)Pixel ratio (%)Median rendered-depth error (mm, 2k → 6k)Anchor coverage (%, 2k → 6k)
(1) [0.5, 1)24.50.71 → 0.3960.4 → 64.9
(2) [1, 2)17.91.35 → 0.6430.8 → 45.2
(3) [2, 4)8.72.61 → 1.128.0 → 22.7
(4) [4, 8)3.55.28 → 1.984.0 → 11.3
(5) ≥8\geq 82.010.69 → 2.192.4 → 7.5

Thank you for the detailed clarifications and extensive additional experiments. The controlled comparisons clarify the respective contributions of anchor construction and the integration framework, and the initialization experiments address my concerns about robustness. My concerns have been resolved, and I will raise my rating to Weak Accept.

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

This paper studies sparse-view surface reconstruction with 3DGS. The key observation is that aligned depth from feed-forward priors such as VGGT has a smoothly varying absolute bias, while its local depth differences are accurate and degrade gradually with pixel distance. The proposed method integrates sparse absolute depth anchors from multi-view photometric matching with VGGT log-depth gradients via gradient-domain integration to produce dense depth supervision, further improved by prior-based cost aggregation and anchor refinement. Experiments on DTU and TnT datasets show improved Chamfer distance and F-score over recent explicit and implicit surface reconstruction methods.

Critical strengths
  1. Clear motivation. The authors observe that aligned VGGT depth has a biased absolute scale but accurate local depth differences, and accordingly propose a dual-evidence integration that combines photometric anchors with VGGT log-depth gradients, together with prior-based aggregation and anchor depth refinement to improve anchor quality.
  2. Sufficient ablations. The ablations clearly demonstrate the effectiveness of the integration formulation and of the PBA/ADR components, and Tab. 11 shows the method generalizes to other feed-forward priors (DA3, MoGe-3, π³).
  3. SOTA performance. Under the same MASt3R-SfM initialization, DGS reduces the DTU Chamfer distance from 0.85 (DP-GS) to 0.71 and improves the TnT F-score from 0.266 to 0.310
  4. The paper is well written and well organized. The motivation, formulation and evidence construction are easy to follow. The appendix provides full details (IRLS, thresholds, reliability criterion) for reproduction, along with more qualitative comparisons.
Critical weaknesses
  1. Limited novelty. The core insight (priors are reliable in relative but not absolute depth) is well known, and the solution, gradient-domain integration with sparse control points, is classical (Horovitz & Kiryati, 2004; Nehab et al., 2005). PBA and ADR also adapt existing MVS techniques (GoMVS, ADP-MVS). The main contribution is combining them for 3DGS supervision. Moreover, DP-GS (Han et al., 2026), which also propagates depth from high-confidence regions with geometric priors, is conceptually very close. The authors should clearly discuss the differences and why global integration is preferable to propagation.

  2. Limited evaluation scope. The quantitative evidence is essentially restricted to DTU. The TnT evaluation uses only 4 scenes and 3 baselines. Since the method relies on photometric matching to obtain anchors, its behavior on textureless, reflective, or large-scale indoor scenes (e.g., ScanNet, Replica, Mip-NeRF 360), where reliable anchors become sparse, is not examined, and no failure cases are shown.

  3. No analysis across sparsity levels. All experiments use a fixed number of views (3 on DTU, 10 on TnT). It is unclear how DGS compares with the baselines as the view count varies: with fewer views, both the VGGT prior and the photometric anchors degrade, and it is unknown whether the integration still helps; with more views, the advantage over direct depth supervision may vanish.

Questions for authors
  1. Could the authors clarify the key difference between DGS and DP-GS (Han et al., 2026), which also propagates depth from high-confidence regions using geometric priors? Given that gradient-domain integration with control points and prior-guided cost aggregation are both established techniques, what do the authors consider the essential technical novelty beyond combining them for 3DGS supervision?

  2. How does the method behave when the view sparsity increases (e.g., 2 views) or when reliable anchors become scarce, such as on textureless, reflective, or outdoor large-scale scenes where photometric matching and the VGGT prior both degrade? Does the integration still help over direct depth supervision, or does it amplify errors in these settings?

3 comments
AuthorsOpenReview ↗

Response to Reviewer bDyW (Part 1)

We sincerely thank the reviewer for the deep, detailed, and constructive feedback. We address each question below and will update the results in the final paper.

Q1: Relation to DP-GS

The key difference lies in how the dense depth is formed from the reliable depths and the prior: DP-GS propagates depth locally and iteratively, whereas DGS solves for all depths jointly through a global integration.

DP-GS propagates depth by iteratively averaging the plane-extrapolated depths of neighboring pixels, with planes defined by the prior normals. In each iteration, every pixel, including the anchors themselves, is re-estimated as this average, and the newly reached pixels are marked as reliable for the next iteration. This leads to two drawbacks. First, when a non-anchor pixel is updated, depths propagated from anchors along short and long paths are weighted equally, although the latter are less reliable, which dilutes the influence of nearby anchors. Second, the repeated averaging acts as a low-pass filter that over-smooths the reliable high-frequency details of the prior. In contrast, integration constrains the absolute depth at the anchors, preserves the local log-depth differences of the prior elsewhere, and solves for all depths in a single objective. It thus keeps the high-frequency details of the prior while correcting part of its errors with the accurate anchors.

To verify this, we compare the two schemes within our framework on DTU, changing only how the supervision depth is generated. The NCC loss and multi-view consistency loss used in DP-GS are not included. We pair each type of anchors (DP-GS and ours) with each densification scheme (i.e., the way of generating dense depth from the anchors and the prior), namely DP-GS propagation and our integration, and report the normal accuracy of the supervision depth and the final CD (Tab. R1). For fairness, propagation uses the same VGGT normals as our method.

  1. Integration outperforms propagation. Propagation increases the mean normal error compared with the VGGT depth, since repeated averaging over-smooths the details of the prior. In contrast, integration preserves these details, and with our anchors it further reduces the normal error from 14.27° to 13.58°. Integration also yields a lower CD under both types of anchors (0.906 vs. 1.255 and 0.707 vs. 0.988).
  2. Our anchors outperform DP-GS anchors under both densification schemes (0.988 vs. 1.255 and 0.707 vs. 0.906).

Tab. R1: Comparison of propagation and integration under different anchors on DTU.

AnchorsDepthNormal error (°) ↓Normal < 5° (%) ↑Normal < 15° (%) ↑CD ↓
–VGGT depth14.2722.669.3–
DP-GSPropagation16.0520.565.61.255
DP-GSIntegration14.2423.769.40.906
DGSPropagation15.6522.867.40.988
DGSIntegration13.5826.771.90.707

Q2: On the contribution

Geometric priors from pretrained models have become a common ingredient in recent NeRF- and 3DGS-based reconstruction. Existing methods (e.g., MonoSDF, DN-Splatter, PlanarGS, VGGS) typically use the prior's depth as absolute supervision. Our main contribution is a more effective way of exploiting such priors, grounded in an analysis of their reliability. The analysis shows that the prior's absolute depth introduces erroneous supervision where reliable photometric evidence exists (Fig. 6a), whereas its relative depth remains reliable locally (Fig. 1). This motivates using the prior only as local differential evidence. Since this relative depth is reliable only over a short range, the integration requires dense and accurate anchors, which is challenging under sparse views. We therefore design PBA and ADR specifically for this purpose, and as shown in Q1, the two parts are complementary. This way of using priors outperforms direct depth supervision (Tab. 4), and the benefit holds across different priors (Tab. 11) and in the dense-view setting (Tab. R3). We hope this perspective can benefit future optimization-based methods that incorporate geometric foundation priors.

AuthorsOpenReview ↗

Response to Reviewer bDyW (Part 2)

Q3: Analysis across sparsity levels

We evaluate DGS and VGGS on DTU under different sparsity levels, with 2, 3, 5, 10, and 15 input views. As shown in Tab. R2, DGS achieves better results across all sparsity levels. With only 2 views, the performance of both methods drops due to the very limited observations, while DGS still outperforms VGGS. With 10 or more views, the gap narrows, since both methods benefit from stronger multi-view constraints, but it does not vanish.

We further evaluate the dense-view setting, where the depth supervision of VGGS and DGS is integrated into 2DGS as a plug-and-play module, which also examines whether it transfers to a different backbone. Both variants share the same 2DGS backbone and training schedule, differing only in the depth supervision. As shown in Tab. R3, DGS reduces the mean CD of 2DGS by 0.20 mm (0.838 → 0.636), compared with 0.09 mm for VGGS (0.838 → 0.748). Moreover, the depth supervision of VGGS degrades the reconstruction on scans 24, 40, 63, and 114, whereas DGS improves the reconstruction on all scenes.

Tab. R2: Chamfer Distance (↓) on DTU with different numbers of input views.

Method2 views3 views5 views10 views15 views
DGS1.2520.7070.7040.6880.680
VGGS1.4340.9130.9100.8240.750

Tab. R3: Chamfer Distance (↓\downarrow) on DTU with dense views.

MethodMean24374055636569
2DGS0.8380.4990.8130.4030.4241.6040.9450.828
2DGS + VGGS0.7480.5100.8060.5060.3971.8300.6780.709
2DGS + DGS0.6360.4430.7740.3260.4100.7810.7090.650
Method8397105106110114118122
2DGS1.3071.0980.6250.6611.7680.4120.6660.511
2DGS + VGGS1.1921.0580.5900.5481.0280.4150.4980.459
2DGS + DGS1.2090.9370.6050.5100.8710.3850.5010.430

Q4: Behavior under scarce anchors

We simulate scarce anchors on DTU by randomly removing anchors in two ways. Pixel-wise removal drops each anchor independently at random, which reduces the anchor density while keeping the anchors spread across the image. Block-wise removal divides the image into non-overlapping 16×16 blocks and drops all anchors within randomly selected blocks, which reduces both the density and the spatial coverage of anchors, leaving large anchor-free regions. In both cases, we retain 50%, 25%, 5%, and 1% of the anchors. On average, the foreground of each view contains about 100k anchors originally and about 1k anchors at the 1% ratio.

As shown in Tab. R4, the integration benefits from both dense and well-spread anchors. As the anchors become much sparser or less spread, its performance gradually degrades toward that of alignment, and becomes slightly worse than alignment at the 1% ratio. In comparison, alignment remains stable across all ratios, since estimating a global scale and shift requires only a small number of reliable anchors. Under the normal setting, DGS provides about 100k anchors per view, which is far above the regime where integration falls behind alignment (around 1%).

Tab. R4: Chamfer Distance (↓\downarrow) on DTU with different ratios of retained anchors.

(a) Pixel-wise removal

Retained anchorsIntegrationAlignment
100% ($\sim$100k)0.7070.907
50%0.7210.912
25%0.7530.909
5%0.7890.907
1% ($\sim$1k)0.9210.911

(b) 16×1616 \times 16 block-wise removal

Retained anchorsIntegrationAlignment
100% ($\sim$100k)0.7070.907
50%0.7450.911
25%0.7780.917
5%0.8120.917
1% ($\sim$1k)0.9500.921
AuthorsOpenReview ↗

Response to Reviewer bDyW (Part 3)

Q5: Evaluation scope

Tanks and Temples

We add DP-GS and FatesGS as baselines on Tanks and Temples, using their official code with the same 10-view protocol and MASt3R-SfM initialization. As shown in Tab. R5, DGS outperforms both methods on all four scenes.

Tab. R5: F-score (↑) on Tanks and Temples with 10 input views.

MethodBarnCaterpillarIgnatiusTruckMean
FatesGS0.09420.01490.01640.24810.0934
DP-GS0.16240.09730.10360.34060.1760
DGS (ours)0.23710.13310.35300.51680.3100

Indoor / textureless scenes

To evaluate indoor scenes with large textureless regions, we further test on 8 scenes from ScanNet++ with 5 input views per scene, using the same MASt3R-SfM initialization for all methods. As shown in Tab. R6, DGS achieves the best results on five of the eight scenes. It performs worse on scenes (3), (5), and (7), and the visualizations of these failure cases are provided in Appendix A.3.5. Scenes (5) and (7) contain large textureless and reflective regions, where DGS produces very few anchors that are concentrated in a small number of areas. As a result, its reconstruction falls behind VGGS. In scene (3), the VGGT prediction contains substantial noise, which leads to poor performance of DGS and even causes VGGS to fail.

Tab. R6: F-score (%, ↑\uparrow) with a 2 cm threshold on ScanNet++ with 5 input views.

Method00a231a370 (1)0affa0f99f (2)0b0bb941b7 (3)0c6c7145ba (4)
VGGS58.158.58Failed2.77
DP-GS44.4513.0420.4853.02
DGS (ours)65.6013.7410.4567.00
Method0d8ead0038 (5)49fe768553 (6)7b4cb756d4 (7)7c31bccde5 (8)
VGGS10.6835.672.0854.88
DP-GS16.1246.20Failed55.63
DGS (ours)9.4048.310.5859.62

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback. We are encouraged that the reviewers find the motivation clear and the analysis of the prior insightful (bDyW, b1Ec, cvjP), the improvements consistent and supported by controlled comparisons and ablations (bDyW, b1Ec, cvjP), the supervision strategy generalizable across different geometric priors (bDyW, cvjP), and the paper well written and easy to follow (bDyW).

In response to the comments, we have conducted the following additional experiments.

  1. Controlled comparison of anchor construction and densification schemes, including DP-GS anchors and propagation (Tab. R1, bDyW and Tabs. R1–R2, b1Ec).
  2. Evaluation under different numbers of input views and in the dense-view setting with 2DGS (Tabs. R2–R3, bDyW).
  3. Analysis of the integration under scarce anchors (Tab. R4, bDyW).
  4. Comparison with DP-GS and FatesGS on Tanks and Temples (Tab. R5, bDyW).
  5. Evaluation on indoor textureless scenes from ScanNet++, with failure cases in Appendix A.3.5 (Tab. R6, bDyW).
  6. Sensitivity to initialization with COLMAP point clouds and noise perturbation (Tab. R3, b1Ec).
  7. Recovery from initial rendered-depth errors and the effect of the index-offset threshold η (Tab. R4, b1Ec and Tabs. R1–R2, cvjP).
  8. Anchor statistics on Tanks and Temples and scene-level comparison with aligned prior depth (Tab. R3, cvjP).
  9. Local depth-difference error of the prior with and without de-shifting (Tab. R4, cvjP).
  10. Comparison of prior-based aggregation with index-wise aggregation (Tab. R5, cvjP).

We have also clarified the difference from DP-GS and the source of novelty, the connection between the motivation and the anchor construction, the role of Eq. 9, and the measurement protocol and breakdown of the computational cost, and corrected the value of η in Tab. 8. All these results and clarifications will be incorporated into the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.