acceptodds
Withdrawn from ICLR 2027

AnchorGeo: Confidence-Gated Depth Alignment for Weakly Observed Regions in Gaussian Surface Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstructing accurate surfaces from multi-view images with radiance fields has achieved remarkable progress in recent years. However, prevailing approaches, primarily based on Gaussian Splatting, remain unreliable in weakly observed regions, where the photometric loss is nearly flat and geometry drifts freely: self-consistency regularizers cannot resolve such regions, while external priors applied indiscriminately fight the photometric evidence exactly where it is decisive. In this paper, we introduce AnchorGeo, a confidence-gated depth-alignment framework that commits geometry precisely where observations run out. Our key finding is that the reliability of the rendered depth is detectable per pixel: a crossing in front of the transmittance peak extracts a view-independent surface, whereas one behind the peak is view-dependent by construction. To locate trustworthy geometry, we first propose a Transmittance-based Confidence Gate that separates reliable anchors from weak pixels directly at rendering time. Subsequently, Confidence-Anchored Depth Alignment is designed to align a monocular depth prior to the anchored geometry in closed form on each object segment and to pull only the weak pixels toward the aligned target, routing external cues into exactly the regions the images cannot constrain. Extensive experiments on DTU and Tanks-and-Temples demonstrate state-of-the-art surface quality with the largest gains concentrated in weakly observed regions, and a per-point recovery/damage decomposition verifies that these gains come from recovering weak regions rather than damaging well-observed ones.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Gaussian Splatting, Surface Reconstruction, Monocular Depth Prior, Confidence Gating, Depth Alignment, Weakly Observed Regions

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_kZ3HOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnchorGeo addresses surface reconstruction in weakly observed regions using a Gaussian opacity-field framework built on GGGS. It combines a transmittance-based front-bias gate with texture confidence and multi-view reprojection information to weight external depth supervision. A depth prior is affine-aligned to the current rendered geometry within SAM segments, and the aligned targets are used to regularize reconstruction. The full system also incorporates a normal prior. Experiments report a mean Chamfer distance of 0.45 mm on DTU and a mean F1 score of 0.62 on Tanks and Temples, compared with 0.48 mm and 0.60 for GGGS.

Critical strengths
  1. Relevant problem and a practical design. Selectively using external geometric priors to improve poorly constrained regions is a useful objective for surface reconstruction. Segment-wise calibration and confidence-weighted supervision form a practical approach to integrating these priors.

  2. Positive results on two established benchmarks. The method improves the reported mean scores over GGGS on both DTU and Tanks and Temples. On Tanks and Temples, it matches or improves GGGS on all six evaluated scenes, with a particularly clear improvement on Meetingroom, from 0.42 to 0.50 F1. These results provide evidence that the complete system can benefit challenging scenes.

Critical weaknesses
  1. The method shares the general idea of calibrating external geometric priors with existing approaches. Its novelty appears incremental, although this is not my main concern.

  2. The main concern is the limited improvement over GGGS in both quantitative and qualitative results. Mean DTU Chamfer distance improves from 0.48 to 0.45 mm, and mean Tanks and Temples F1 increases from 0.60 to 0.62. The qualitative differences are also not consistently pronounced. The ablation results show similarly small gains over GGGS, leaving the practical benefit of the proposed components insufficiently convincing.

Questions for authors
  1. What is the main independent contribution beyond geometry-anchored prior calibration and confidence-weighted supervision?

  2. What exactly changes between the gated and ungated configurations in Table 3? Does this comparison remove only the front-bias gate while keeping the remaining alignment and weighting components fixed? Please provide per-scene scores and, where available, repeated-run variability to help assess the reported differences.

  3. Can the authors provide matched close-ups or ground-truth error visualizations against GGGS in the weakly observed regions highlighted by the paper, together with the recovery/damage results claimed in the abstract? These would help assess whether the modest aggregate gains reflect consistent improvements in the intended regions.

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

This paper presents AnchorGeo, a confidence-gated depth alignment framework for Gaussian surface reconstruction in weakly observed regions. The authors investigate the view dependence of transmittance-based surface extraction and propose a confidence criterion to distinguish reliable from unreliable geometric readouts. Based on this analysis, the method aligns monocular depth predictions with reliable geometric anchors within individual segments and selectively applies depth supervision to weakly constrained regions. Experiments on DTU and Tanks-and-Temples demonstrate improvements over the GGGS baseline, achieving a mean Chamfer distance of 0.45 mm on DTU and a mean F1-score of 0.62 on Tanks-and-Temples.

Critical strengths

S1.

The paper provides an interesting analysis of how front-side and back-side transmittance crossings affect the view consistency of surface extraction. The theoretical discussion, together with the supplementary derivations, offers a useful perspective on understanding geometric uncertainty in Gaussian surface reconstruction.

S2.

Rather than applying monocular depth supervision uniformly, the proposed framework aligns depth priors to reliable geometric anchors and focuses supervision on weakly constrained regions. This is a reasonable and practically useful design. The improvements over GGGS on DTU and Tanks-and-Temples provide encouraging evidence of its effectiveness, although further ablations are needed to isolate the contributions of individual components.

Critical weaknesses

W1.

Although AnchorGeo improves the average reconstruction quality over GGGS, the gains are relatively modest and not consistently observed across scenes. On DTU, the mean Chamfer distance decreases from 0.48 mm to 0.45 mm, but 6 out of 15 scenes exhibit regressions (Scans 37, 40, 55, 106, 118, and 122). On Tanks-and-Temples, the mean F1-score increases from 0.60 to 0.62, with several scenes showing only marginal improvements.

More importantly, these regressions raise concerns about the claim that the proposed method preserves reliable geometry while refining weakly observed regions. Although the depth-alignment loss emphasizes weak pixels, it does not strictly prevent updates to reliable geometry. Equation (7) may assign nonzero supervision weights to anchor pixels, and shared Gaussian parameters allow updates driven by weak regions to affect neighboring reliable structures.

While these observations do not establish the causes of the regressions, additional quantitative evidence separating geometric recovery from damage is needed to substantiate the claimed preservation of well-observed structures.

W2.

The theoretical analysis establishes a conditional stability property for transmittance-based surface extraction. However, the practical confidence gate in Eq. (6) relies on the difference between rendered depth and the alpha-weighted mean depth of Gaussian centers.

In scenes containing multiple overlapping Gaussians, this approximation does not necessarily correspond to the theoretical front/back criterion. Consequently, the theoretical stability analysis does not directly guarantee the reliability of the implemented gate.

Although the supplementary material discusses the assumptions and limitations of the analysis, it remains unclear how accurately the practical gate identifies geometrically unreliable regions. A quantitative evaluation relating gate predictions to actual geometric errors, or comparing the implemented gate with the theoretical criterion, would strengthen the paper's central contribution.

W3.

The proposed framework combines texture-based confidence estimation, monocular depth priors, segment-wise affine alignment, and confidence-weighted geometric supervision. While their integration is reasonable, several of these components are based on established techniques.

The main methodological novelty appears to lie in the transmittance-based confidence gate. However, Table 3 primarily compares gated and ungated supervision without sufficiently isolating the contribution of this particular criterion.

Comparisons against texture-only confidence weighting, segment-wise depth alignment without transmittance gating, and alternative confidence estimation strategies would help establish whether the proposed gate offers a distinctive advantage over simpler approaches.

W4.

The supplementary material describes a reliability-weighted TSDF fusion procedure for DTU mesh extraction, incorporating geometric consistency and transmittance-based reliability information.

Since confidence information is involved in both geometric optimization and surface extraction, the reported improvements may reflect contributions from both stages. It is therefore difficult to determine how much of the improvement originates from the optimized Gaussian representation itself.

An ablation comparing standard and reliability-weighted extraction under controlled optimization settings, together with a comparison using matched extraction settings for GGGS and AnchorGeo, would help clarify the contribution of each stage.

W5.

The framework incorporates local texture confidence, multi-view reprojection evidence, and transmittance-based reliability. However, these indicators do not necessarily provide a complete characterization of geometric observability. Low-texture regions can be well constrained by multiple views, while highly textured regions may remain ambiguous under insufficient viewpoint coverage or parallax.

Since improving weakly observed regions is the central motivation, experiments with systematically reduced input views or restricted viewpoint coverage would provide more direct evidence of effectiveness.

The Courthouse analysis in the supplementary material also suggests that substantial geometric ambiguity can remain under limited angular coverage. Further investigation of these failure cases would clarify the method's limitations.

W6.

Table 3 compares ungated and gated normal priors, reporting mean F1-scores of 0.60 and 0.61, respectively. However, Eq. (8) and the implementation details do not explicitly define the normal loss or its gating and weighting mechanism.

Equation (7) defines the depth-alignment loss but does not establish whether normal supervision uses the same texture confidence, multi-view evidence, transmittance criterion, or segment reliability.

Clarifying the normal-prior gating formulation is necessary for reproducibility and for interpreting the corresponding ablation results.

Questions for authors

Q1. Are the DTU regressions caused by misclassifying or unnecessarily updating previously reliable geometry?

Compared with GGGS(the immediate baseline), Chamfer distance worsens on 6 out of 15 DTU scenes. Could the authors use an identical mesh extraction pipeline, classify regions as reliable or unreliable based on the GGGS reconstruction’s error against reference geometry, and separately report recovery, damage, and newly introduced spurious geometry? Please also analyze the texture confidence, front-bias classification, and actual prior-supervision weights in the degraded regions. Were any of these regions already accurate but nevertheless subjected to unnecessary or incorrect supervision? Repeated runs would help distinguish systematic negative transfer from optimization variability.

Q2. What is the exact implementation of the gated-normal configuration in Table 3?

Neither the main paper nor the supplementary material explicitly provides the normal-loss and gating formulas. Please specify the error function, per-pixel weights, and gradient handling for Ln, and clarify exactly what changes between the gated and ungated configurations. Does normal supervision use the transmittance criterion in Eq. (6)? In particular, how are front-side, back-side, and neighboring-view-invisible pixels treated? Please isolate the contribution of the transmittance gate while keeping the other confidence factors fixed.

Reviewer_wTo8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnchorGeo, which addresses geometric drift in weakly observed regions of Gaussian surface reconstructions. The paper first analyzes when a transmittance-based surface readout is consistent across views, and then proposes a confidence gate based on the relative position of rendered depth and contributing Gaussian centers. Confidence-Anchored Depth Alignment is designed to align a monocular depth prior to the anchored geometry in closed form on each object segment and to pull only the weak pixels toward the aligned target. Building upon GGGS, AnchorGeo improves mean DTU Chamfer distance from 0.48 to 0.45 mm and mean Tanks-and-Temples F1 from 0.60 to 0.62. The main contribution is the selective integration of external geometry priors, motivated by an analysis of front- versus back-side transmittance crossings.

Critical strengths
  1. The paper addresses a meaningful failure mode in surface reconstruction. Photometric supervision can leave geometry poorly constrained in textureless or sparsely observed regions. The idea of calibrating an external prior against better-constrained regions is intuitive and makes sense.
  2. The alignment mechanism is simple and computationally efficient. Segment-wise affine fitting accommodates local scale and shift differences in the depth prior. The closed-form 2x2 solution and detached fitting target does not introduce additional learned optimization module.
  3. The ablation on prior supervision supports selective integration. In Table 3, ungated depth supervision reduces mean F1 to 0.58, while the gated version achieves 0.61. This demonstrates that applying a strong prior everywhere can be harmful and that selective supervision matters.
Critical weaknesses
  1. The implemented gate is not a guaranteed theoretical condition. Equation 6 checks whether depth lies before the weighted mean Gaussian-center depth. This does not mean that it precedes every contributing peak. For example, a depth of 2 lies at the mean of centers at 1 and 3 while already passing the first center. It would be better to justify this approximation or present it as a heuristic.
  2. The ablation study does not isolate the proposed transmittance criterion. "Gated” supervision combines texture weighting, segment restrictions, reprojection evidence, fit-quality checks, and transmittance handling. Table 3 does not identify which component causes the improvement. A key missing comparison is the matched pipeline with only the transmittance gate removed or replaced by a simpler confidence measure.
  3. How the gate is used during anchor fitting is not very clear. Equation 5 says that fitting weights favor high-confidence anchors, but does not define those weights or explicitly show how back-side anchors are excluded or discounted. Equation 7 instead uses the transmittance gate to reset reprojection-error weights. It would be better to specify how the gate affects anchor eligibility, fitting weights, and prior-loss weights.
Questions for authors
  1. How does the geometric error distribution for accepted and rejected pixels look like? How often does the mean-depth proxy disagree with a contribution-level test?
  2. What are the fitting weights in Equation 5 and does transmittance confidence affect them directly?
  3. What happens when the full method retains texture weighting, SAM segments, and affine fitting but removes only the transmittance gate?
  4. Do the reported times include SAM, depth-prior, and normal-prior inference? What is the total preprocessing and reconstruction cost?
Reviewer_knAUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper targets weakly observed regions in GS surface reconstruction. Reconstruction of pixels whose rendered depth is unreliable is guided by Depth Anything 3, aligned per SAM segment. The results demonstrated small gains over the GGGS baseline on DTU and Tanks-and-Temples.

Critical strengths

The main observation and motivation are interesting. The authors show that when a camera ray is half blocked on the far side of a Gaussian, different cameras see different surfaces. They use this view inconsistency to flag unreliable pixels. This is verified on DTU: regions with such pixels have 2–3× higher error.

Critical weaknesses
  • Relying on a monocular depth prior is hard to justify for this problem. The task is multi-view reconstruction, yet the fix for weak regions is single-view depth, which is itself inaccurate and inconsistent across views. This conflicts with the paper’s own diagnosis: the problem is that different views see different surfaces, and the per-view depth targets have the same issue.
Questions for authors
  1. Why a monocular prior, not a multi-view consistent prior?
  2. For a same weak region, will the DA3's depth estimations from different views agree in 3D? Please measure how far apart they are. How to solve this same fundamental inconsistency issue?

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

Related papers

Discussion (0)

Sign in to comment.