- 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
-
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.
-
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.
-
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.
-
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
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.
- 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.
- 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.
| Scene | Anchor coverage (%, 2k → 6k) | Mean distance to nearest anchor (px) | P90 distance (px) | < 5 px (%) | < 10 px (%) | < 15 px (%) | F-score, aligned prior | F-score, DGS |
|---|---|---|---|---|---|---|---|---|
| Barn | 12.5 → 25.9 | 8.2 | 22.1 | 53.9 | 70.4 | 80.3 | 0.1518 | 0.2371 (+0.0853) |
| Caterpillar | 9.9 → 19.3 | 8.1 | 19.0 | 46.8 | 67.8 | 81.3 | 0.0918 | 0.1331 (+0.0413) |
| Ignatius | 56.2 → 62.4 | 2.8 | 9.0 | 82.5 | 90.9 | 94.8 | 0.2254 | 0.3530 (+0.1276) |
| Truck | 23.4 → 33.9 | 4.5 | 11.7 | 64.3 | 84.1 | 93.7 | 0.3228 | 0.5168 (+0.1940) |
| Mean | – | – | – | – | – | – | 0.1980 | 0.3100 (+0.1120) |
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 distance | 1 px | 2 px | 5 px | 10 px | 20 px | 50 px | 100 px |
|---|---|---|---|---|---|---|---|
| Scale alignment | 0.08 | 0.13 | 0.24 | 0.37 | 0.53 | 0.89 | 1.25 |
| Scale-shift alignment (Fig. 1) | 0.08 | 0.13 | 0.22 | 0.33 | 0.48 | 0.75 | 0.93 |
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.
| Aggregation | Anchor coverage (%) | Anchors with error < 2 mm (%) | CD ↓ |
|---|---|---|---|
| None (per-pixel) | 49.6 | 89.1 | 0.834 |
| Index-wise aggregation | 55.0 | 91.6 | 0.748 |
| PBA | 57.8 | 93.6 | 0.707 |
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.
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.
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.
Tab. R2: Effect of the index-offset threshold η on DTU.