acceptodds
Under review as a conference paper at ICLR 2027

UrbanBP-GS: Building-Aware Partitioning and Planarity-Constrained Updating for Morphology-Adaptive Urban Novel-View Synthesis

Abstract

Block-wise reconstruction based on 2D Gaussian Splatting (2DGS) has emerged as a promising paradigm for high-fidelity novel-view synthesis of large urban scenes. However, this paradigm encounters bottlenecks in scene partitioning and primitive representation. Partitioning schemes suffer from inter-block primitive imbalance and building fragmentation, while 2DGS limits the rendering of geometrically irregular structures. To address these limitations, we propose UrbanBP-GS, a morphology-adaptive framework that uses building-aware scene partitioning to balance primitive counts and preserve building connectivity, and planarity-constrained 3D Gaussian updating to support high-fidelity urban rendering with irregularity-tolerant primitives. For building-aware scene partitioning, UrbanBP-GS first initializes scene partitioning under a primitive-balance constraint, then detects connected buildings spanning multiple blocks and consolidates each building into its dominant block under a per-block primitive-count variation budget. For planarity-constrained 3D Gaussian updating, UrbanBP-GS associates each planar region with its contributing Gaussians, estimates their normals through soft shortest-axis selection, and promotes orientation consistency among Gaussians in the same planar region while preserving nonzero-thickness primitives to represent irregular structures. Extensive experiments demonstrate that UrbanBP-GS achieves superior PSNR and LPIPS performance across multiple large-scale urban scenes, with particularly notable gains in densely built areas.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Urban Novel-View Synthesis, Building-Aware Partitioning, Plane-Guided Optimization

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 5 reviews, above 27% of the others.
Reviewer_ut9POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper improves block-wise urban novel-view synthesis built on 2D Gaussian Splatting per block. It identifies two bottlenecks: partitioning can imbalance primitive counts across blocks or split connected buildings across independently optimized blocks, and zero-thickness 2D primitives fit planar structure well but are too rigid for irregular geometry such as vegetation or cranes. UrbanBP-GS contributes building-aware partitioning, which starts from a primitive-balanced partition and reassigns whole cross-block buildings to their dominant block under a variation budget, and planarity-constrained updating, which keeps full 3D Gaussians but aligns the orientations of those contributing to the same detected plane. Experiments on four urban scenes against Gaussian-splatting baselines show the best perceptual quality on all scenes and the best PSNR on three, with remove-one ablations on a single scene.

Critical strengths
  1. Building-aware partitioning targets a concrete failure mode of block-wise pipelines, and budget-constrained reassignment is a simple, controllable fix that preserves the load-balance purpose of the initial partition. The formulation is precise about what the budget does and does not constrain, and one ownership map consistently drives view assignment and merging, which makes the method easy to re-implement. The results are consistent with the motivation: perceptual quality is the best of all methods on every scene, with clear margins over the strongest baselines, and the reporting concedes the scene where PSNR trails a baseline.
  2. Keeping finite-thickness primitives while enforcing orientation consistency within planes is a reasonable middle ground between the surface bias of 2D Gaussians and the flexibility of 3D ones, and the planarity module avoids two standard pitfalls: the soft axis selection stops gradients that would merely switch axes, and the alignment loss treats opposite normals as equivalent. Qualitative comparisons on thin and irregular structures support the design.
Critical weaknesses
  1. Attribution is never isolated. The claim that zero thickness is the limiting factor is untested: the ablation varies the proposed components but includes no primitive swap, and the main comparison is confounded because baselines differ simultaneously in primitive, partitioning, regularization, and inherited depth supervision. Notably, the 2DGS representative of the same lineage scores below the plain 3DGS variant of that lineage on most scenes and metrics, and the paper does not state whether that row was re-run under the present protocol or quoted; if quoted, the comparison to it is not meaningful. As it stands, no reported margin can be assigned to any specific component.
  2. The two central claims are never quantified, and the geometric mechanism is left partly unspecified. The formulation defines per-block primitive counts and the target balance, yet no counts, imbalance ratios, or fragmentation statistics are reported before or after refinement, and no seam or consistency measure accompanies the stated goal, although this is the paper's central motivation. On the geometric side, only photometric novel-view metrics are reported despite the planarity motivation; plane identities are not matched across views, so the same facade is aligned toward independent orientations in different views, and sensitivity to plane-detection errors is unexamined; the shaping loss is described only in prose; and no training or rendering cost, loss weights, or limitations are given despite the large-scale framing.
Questions for authors
  1. Can you add a primitive-swap ablation — same partitioning and losses, zero-thickness or 2DGS primitives — across the scenes, and state whether the 2DGS baseline row was re-run under your settings or quoted from elsewhere?
  2. What are the per-block primitive counts before and after refinement, how many buildings span multiple blocks initially and finally, and how sensitive are the results to the variation budget and the shaping threshold? What is the exact functional form of the shaping loss?
  3. Do the baselines receive the same depth supervision and resolution policy, the same train-test split and test-view count, and were they retrained under your protocol? How much of the improvement survives when only supervision is equalized?
  4. How are per-view plane orientations reconciled for the same facade, and how sensitive is the method to plane-mask errors? Can a geometric evaluation — depth or normal error, or surface quality on facades — be added to test the geometric claims directly?
4 comments
AuthorsOpenReview ↗

Attribution and Primitive Choice

Response to Reviewer (1/3) – Attribution and Primitive Choice

We thank the reviewer for the constructive evaluation and for recognizing the practical value of building-aware partitioning and finite-thickness Gaussians with planar orientation consistency. We provide additional measurements and clarify the scope of our experimental comparisons.

W1 / Q1 / Q3 – Attribution under matched supervision.

We distinguish component attribution within UrbanBP-GS from cross-paper comparisons. In our controlled ablations, all variants share the same monocular inverse-depth supervision, image resolution, train–test split, coarse checkpoint, number of blocks, and 60k steps per block. In addition to the Residence ablation (Table 2), we evaluated the components on SMBU and included a Base that removes building-aware refinement, Gaussian shaping, and planarity consistency together:

Variant (SMBU)PSNR ↑SSIM ↑LPIPS ↓
Full24.9020.8090.171
w/o building-aware24.8620.8070.174
w/o planarity24.9020.8080.173
w/o shaping24.8350.8070.173
Base (all removed)24.8020.7960.181

With supervision and training iterations matched, removing all three components decreases PSNR by 0.100 dB and SSIM by 0.013, while increasing LPIPS by 0.010. Every individual removal also increases LPIPS on both Residence and SMBU. The planarity term shows a modest LPIPS benefit but no distinguishable PSNR gain on SMBU at the reported precision.

The joint SSIM and LPIPS degradations exceed the sums of the individual remove-one degradations, suggesting complementary or non-additive interactions among the components. These results isolate their contributions within our common backbone, rather than attributing the full margins over external baselines to individual modules.

Primitive choice and CityGaussianV2 provenance.

We clarify that UrbanBP-GS retains full 3D Gaussians during block-wise optimization; it does not use zero-thickness 2D primitives. Our motivation is to retain flexibility for irregular geometry while applying orientation constraints selectively to detected planar regions.

Table 1 quotes baseline results from published sources rather than independently retraining those methods. Specifically, the CityGaussianV2 entries for Mill19 and UrbanScene3D are taken from the MetroGS (CVPR 2026) supplementary evaluation, since the original CityGaussianV2 paper does not report these scenes. We will explicitly identify the source of each row.

Our evaluation follows the stated benchmark resolution policy (4x downsampling for Mill19 and UrbanScene3D; 1.2x for MatrixCity-Aerial) and the corresponding train–test splits. However, external methods may differ in depth supervision, training budgets, block configurations, and other implementation details. Thus, matching the benchmark protocol does not make their training conditions identical. We do not interpret cross-paper margins as controlled evidence of any single component's contribution.

Regarding primitive choice, CityGaussianV2 reports higher geometric F1 than CityGaussian on GauU-Scene (0.501 vs. 0.453), while obtaining lower rendering quality on the three reported scenes. For example, on Russian Building, CityGaussianV2 reports PSNR/SSIM/LPIPS of 24.12/0.784/0.196, versus 24.37/0.808/0.163 for CityGaussian. This illustrates a possible geometry–rendering tradeoff, but is not a primitive-only comparison because the methods also differ in optimization.

We therefore present retaining the third Gaussian scale as a design choice motivated by the coexistence of planar and irregular urban structures, rather than claiming that zero thickness alone explains the performance differences. Our controlled ablations support the benefits of the proposed updating and partitioning components within this 3DGS representation.

AuthorsOpenReview ↗

Partition Statistics and Loss Formulation

Response to Reviewer (2/3) – Partitioning Measurements and Loss Formulation

W2 / Q2 – Quantifying building integrity and workload balance.

We measured fragmentation and per-block primitive counts on SMBU (103 extracted candidate regions, 4 blocks) under different refinement budgets:

BudgetCross-block regionsFragmented areaPer-block primitives (M)Max/meanMin/mean
Initial partition1845.7%2.28 / 2.31 / 2.27 / 2.221.0170.977
0.1–0.5 (default: 0.3)127.3%2.50 / 2.37 / 2.03 / 2.171.1010.896
At least 0.6600.0%–1.5960.480

Structural integrity. Refinement consolidates 17 of the 18 initially cross-block candidate regions. The remaining region corresponds to a large vegetated hillside included in the extracted candidates, rather than an ordinary building instance. It accounts for the remaining fragmented candidate-region area.

Budget sensitivity and balance. All tested budgets between 0.1 and 0.5 produce the same partition. Although the budget limits each reassignment rather than cumulative deviation, the largest cumulative block-count change is only 10.5%. Consolidating the remaining large region requires a budget of at least 0.66 and would increase the maximum block load to approximately 1.60 times the mean. Thus, the budget rejects a disproportionately expensive reassignment.

The trained SMBU model contains 13.2M Gaussians after densification, distributed as 3.64 / 3.77 / 2.81 / 3.00 M across the refined blocks. The maximum load ratio increases only from 1.101 to 1.142 after training, indicating that the observed imbalance remains moderate.

Rendering contribution. Relative to the initial partition used by the w/o building-aware variant, refinement improves PSNR/SSIM/LPIPS from 24.862/0.807/0.174 to 24.902/0.809/0.171. The fragmentation measurements directly quantify structural consolidation and identify where block boundaries intersect extracted regions. They do not themselves measure image-space seam error, but provide complementary evidence for the mechanism underlying the rendering improvements.

Exact shaping loss.

Let ρi=smin⁡,i/smax⁡,i\rho_i=s_{\min,i}/s_{\max,i} and define the active set

S=i∣ρi>τs.S=\\{i\mid\rho_i>\tau_s\\}.

The shaping objective is

\qquad \tau_s=0.2.$$ Only Gaussians whose shortest-to-longest scale ratio exceeds the threshold are penalized. Once $\rho_i\leq\tau_s$, the Gaussian leaves the active set. This encourages flattening without enforcing zero thickness. The threshold controls which Gaussians receive shaping supervision; we use the fixed value 0.2 in the reported experiments and do not claim robustness to untested threshold values. **Loss settings.** The complete objective is $$L=L_{\mathrm{rgb}}+w_dL_{\mathrm{depth}}+w_sL_{\mathrm{shape}}+w_pL_{\mathrm{planarity}},$$ where $$L_{\mathrm{rgb}}=0.8L_1+0.2(1-\mathrm{SSIM}).$$ The depth loss applies the same L1/SSIM combination to Depth-Anything-V2 inverse depth. With $$h(t)=\min(t/60000,1),$$ the loss weights are $$w_d=0.5(0.005)^{h(t)},$$ $$w_s=0.01(0.01)^{h(t)},\quad t\geq3000,$$ $$w_p=0.2(0.05)^{h(t)},\quad t\geq6000.$$ The shaping and planarity weights are zero before their respective activation steps. The selector temperature is $T=0.1$. The coarse stage (30k steps) uses only photometric and depth supervision. These settings will be stated explicitly in the appendix.
AuthorsOpenReview ↗

Plane Consistency, Robustness and Cost

Response to Reviewer (3/3) – Plane Consistency, Robustness, and Computational Cost

Q4 – Cross-view plane consistency and mask errors.

Cross-view orientation coupling. Plane masks determine which Gaussians are grouped within each training view; they do not provide externally estimated target normals. Each group's reference direction is computed from the current orientations of its contributing Gaussians.

When the same facade is observed across multiple views, the associated groups can share many of the same underlying 3D Gaussians. Because these Gaussians have persistent parameters across views, their orientations receive coupled supervision rather than being independently optimized for each image.

We acknowledge that this is implicit coupling, not explicit cross-view plane-instance matching. Therefore, the method does not guarantee identical group directions across all observations of a facade. Nevertheless, the trained SMBU model exhibits coherent per-instance orientations: over 100 training views and 869 detected plane instances, the per-view median resultant length of the weighted signed normals ranges from 0.88 to 0.94.

Mask-error handling. Contributor collection applies visibility and opacity filtering in the rasterizer. A Gaussian is recorded only when its per-pixel opacity satisfies α≥1/255\alpha\geq1/255 and ray transmittance remains at least 10−410^{-4}. We retain the 10 largest plane instances per view, discard masks smaller than 10 pixels, deduplicate contributor IDs, and exclude instances with fewer than 3 Gaussians.

Associations are formed per view without persistent global plane labels. Each plane-instance loss is normalized by the total weight of its associated Gaussians. Thus, imperfect masks do not impose permanent hard plane assignments, although erroneous associations may still influence subsequent optimization through parameter updates.

The planarity weight decreases from approximately 0.148 at activation to 0.01 at the end of training, reducing the relative influence of plane supervision during later iterations. These mechanisms mitigate reliance on noisy associations but do not guarantee immunity to plane-detection errors.

Geometric evidence and limitations. We distinguish geometric consistency from ground-truth geometric accuracy. The orientation-resultant statistics characterize within-plane Gaussian alignment, while the partition measurements directly quantify structural fragmentation. Improvements in held-out-view SSIM and LPIPS provide complementary rendering evidence, but do not directly measure surface-normal or depth accuracy.

Our current evaluation therefore supports structural consolidation, orientation coherence, and improved novel-view synthesis, without establishing a quantitative gain in ground-truth geometric accuracy. We will make this distinction explicit in the limitations discussion.

Computational cost.

Measurements on SMBU (500 training images, 1600x1064) use four RTX 4080 SUPER GPUs modified to 32 GB memory each:

ItemCost
ZeroPlane plane masks5.2 min
Depth-Anything-V2 inverse depth5.5 min
Coarse training (30k steps)1.37 h
Building extraction, partition, view assignment2.0 min
Iteration time, planarity off to on252 to 423 ms
Peak allocated GPU memory, off to on17.9 to 20.8 GB
Total reconstruction, including preprocessingApproximately 9 h on 4 GPUs
Final model13.2M Gaussians, 23.1 FPS at 1600x1064

The iteration-time and memory measurements are taken on the largest block model (10.4M Gaussians before ownership filtering). Contributor tracking and plane association increase training-time computation and memory, but are not required during rendering.

Plane-mask prediction, depth estimation, and geometry processing together take approximately 13 minutes, less than 3% of the total reconstruction time. The four blocks are trained in parallel, one per GPU.

We will explicitly discuss the remaining limitations: the lack of a controlled primitive-swap comparison, dependence on plane-mask quality, absence of direct geometric ground truth evaluation, and additional training-time overhead.

We thank the reviewer again for the constructive suggestions. We believe these measurements and clarifications better establish the specific contributions, empirical evidence, and limitations of UrbanBP-GS.

My assessment rests on three points. 1) Every baseline result in Table 1 is quoted from published sources rather than independently retrained under this paper's own depth supervision and training budget. 2) And decisively, the questions that would decide acceptance are answered with acknowledged limitations rather than with experiments. 3) The SMBU ablation shows that removing all three components together decreases PSNR by about 0.1 dB, that the planarity term alone brings no distinguishable PSNR gain at the reported precision, and that the building-aware refinement yields only a 0.04 dB improvement. My assessment is unchanged, and my score remains 2.

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

This paper proposes a framework for the novel-view synthesis of large urban scenes based on block-wise 3D Gaussian updating. To overcome the building fragmentation issue as well as the imbalance budgets, building-aware partitioning is introduced, which iteratively refines block assignments to avoid dividing a building into multiple blocks while maintaining primitive-count variation to satisfy the predefined budget. Another improvement is the planarity-constrained 3D Gaussian updating, which first uses threshold-controlled operation to build a local short axis with three optimizable scale parameters, uses soft surface-direction estimation to convert the axes into a continuous surface direction with weights, and then groups the visible Gaussians into each planar region and aligns the orientation with the shared orientation for each group. The experiments demonstrate the proposed method achieves the best LPIPS across 4 urban scenes and remove-one ablations show incremental improvements of model performance by single modules within the full framework on residential scenes.

Critical strengths
  1. The building-aware partitioning is well motivated by the clearly explained problem of buildings being fragmented across training blocks. The proposed partitioning method takes primitive-count variation into account and reduces building fragmentation with simple, intuitive reassignment criteria.
  2. The proposed soft surface-direction estimation and planar orientation alignment together provide a feasible way to use plane priors to provide additional geometric supervision.
  3. The empirical experiments show that the framework achieves the best LPIPS across all 4 urban scenes and the best PSNR across 3 out of 4 urban scenes.
Critical weaknesses
  1. Eqs. 3–4 raise sign-consistency concerns. Equivalent axis-sign choices can change the soft surface direction in Eq.3 while keeping the Gaussian covariance unchanged. Eq.4 computes its reference group direction with the signed mean, which is not invariant to individual normal flips and has no specified direction under exact cancellation. The squared term here does not resolve these issues. The paper should clarify whether a mechanism is used to addresses these ambiguities.
  2. The ablation is insufficient.
    • Remove-one experiments are reported only on Residence scene but no other 3 scenes.
    • The improvement of every single component seems some kind of modest, like LPIPS is 0.184 vs 0.187 vs 0.185. These differences are much smaller than the gap to the second best LPIPS result on Residence, LoBE-GS at 0.209. Besides remove-one ablation, we need a matched baseline without the proposed modules and also joint-removal ablations.
  3. The partitioning benefits and computational tradeoff are not evaluated quantitatively. This paper claims to be able to reduce building fragmentation while keeping the primitive-count changes of accepted refinements controlled. However, there is no sensitivity analysis of budget δ on model performance.
Questions for authors
  1. Is there a consistent sign convention on the local Gaussian axes imposed before computing the soft direction in Eq.3? When flipping the signs of two principal axes, the Gaussian covariance will keep unchanged but their weighted sum will change. In this case, equivalent representations of the same ellipsoid can produce different estimated normal orientations. How the proposed framework handle this ambiguity?
  2. What is the exact matched baseline with all proposed modules disabled? How does its performance compare with other methods in the paper?
  3. How was the budget δ chosen? Would the budget δ allow the number of fragments to be reduced to nearly zero or still leave some fragments just as shown Figure 3? If there is still a high fraction of building fragments, how large the budget δ should be to ensure almost no fragments? Is such a budget affordable?
9 comments

Response to Reviewer (1/2)

We thank the reviewer for the detailed feedback, and for recognizing that preserving building continuity while controlling block sizes is a useful consideration and that combining structure-aware partitioning with planar regularization is intuitively reasonable. We respond below with new measurements and will incorporate all of them in the final version.

W1 – Presentation. We will restructure the introduction around a single question: which structures must remain intact within an independently optimized block, and which primitive represents them faithfully. The partitioning argument (L065) will then lead directly into the primitive argument (L071). Eqs. (3)–(4) will be split into separately numbered definitions, Fig. 2 annotations enlarged, Fig. 3 replaced by a high-contrast difference map, and repetitive descriptions condensed.

W2 – Algorithmic differences from PG-SAG and PGSR.

(i) Partitioning vs. PG-SAG. PG-SAG groups the scene into building-level sub-regions using semantic masks from a cross-modal segmentation model, then reconstructs each group in parallel. Our refinement needs no semantic model. It operates directly on primitive ownership in BEV: connected building regions are extracted from the coarse geometry, and each region is merged into the block it overlaps most, only if every affected block's relative primitive-count change stays within a budget δ\delta. This makes the fragmentation–load trade-off explicit while constraining each reassignment. On SMBU, buildings crossing a block boundary drop from 18 to 1, while the largest cumulative block change is 10.5% (Q3).

(ii) Planar regularization vs. PGSR. PGSR promotes planar Gaussians by minimizing their smallest scale and enforces single-view depth–normal and multi-view geometric consistency for surface reconstruction. We instead (a) keep finite-thickness 3D Gaussians: the shaping term stops once a primitive reaches a bounded aspect ratio (ρ≤τs\rho \le \tau_s) instead of driving its smallest scale to zero, so vegetation and terrain retain non-zero thickness; and (b) impose orientation consistency only among Gaussians that actually contribute to pixels of the same detected plane instance, identified through per-pixel contributor IDs from an extended rasterizer. Regularization therefore follows instance-level planar structure rather than applying uniformly to all primitives.

(iii) Finite thickness vs. 2DGS. We keep finite-thickness Gaussians because published results show that the 2DGS route mainly benefits geometry rather than rendering. 2DGS substantially improves surface reconstruction while reporting appearance quality competitive with 3DGS. CityGaussianV2, the 2DGS-based successor of CityGaussian, shows the same trade-off in its own evaluation, where both methods are evaluated under one protocol: on GauU-Scene it reports higher geometric F1 than CityGaussian (0.501 vs. 0.449) but lower PSNR and SSIM and higher LPIPS on all three scenes (e.g., Russian Building: 24.12 / 0.784 / 0.196 vs. 24.37 / 0.808 / 0.163). Since urban scenes also contain substantial non-planar content such as vegetation and terrain, we retain the third scale for novel-view synthesis and add planarity only where planes are detected.

W3 – Baselines. We will add Momentum-GS, MixGS, SigGS, CityGS-X, UrbanGS and MetroGS to Table 1. Numbers are taken from each paper (PSNR / SSIM / LPIPS); protocol differences, such as MetroGS's per-image appearance embeddings, will be listed in the caption.

MethodBuildingRubbleResidenceMatrixCity-Aerial
Momentum-GS23.65 / 0.813 / 0.19426.66 / 0.826 / 0.20023.37 / 0.828 / 0.19629.11 / 0.881 / 0.180
MixGS23.03 / 0.771 / 0.26126.66 / 0.792 / 0.26723.39 / 0.815 / 0.219–
SigGS (SIG-L)22.86 / 0.778 / 0.22527.35 / 0.843 / 0.18923.46 / 0.840 / 0.19429.04 / 0.888 / 0.184
CityGS-X22.76 / 0.817 / 0.19126.15 / 0.823 / 0.21022.44 / 0.819 / 0.19427.58 / – / –
UrbanGS22.82 / 0.802 / 0.20826.25 / 0.791 / 0.21022.48 / 0.823 / 0.205–
MetroGS23.06 / 0.787 / 0.17327.48 / 0.826 / 0.14723.38 / 0.824 / 0.16627.52 / 0.854 / 0.167
Ours22.63 / 0.787 / 0.19526.09 / 0.803 / 0.20323.32 / 0.828 / 0.18427.44 / 0.871 / 0.139

UrbanBP-GS achieves the lowest LPIPS on MatrixCity-Aerial (0.139 vs. 0.167–0.184) and lower LPIPS than MixGS and UrbanGS on every scene where they report results. Our contributions act on how the scene is partitioned and how primitives are oriented, and are complementary to the representation and training designs of these methods (self-distillation, holistic decoding, signal-aware scheduling, appearance-decoupled 2DGS). PG-SAG and PGSR are discussed in W2 and will be added to Related Work.

Response to Reviewer (2/2)

W4 / Q2 – Ablations. Table 2 removes each component from the full model on Residence while keeping the coarse model, training schedule and evaluation identical. Removing any component lowers PSNR and raises LPIPS: Gaussian shaping costs 0.284 dB, building-aware partitioning 0.165 dB and planarity consistency 0.127 dB, with LPIPS rising from 0.184 to 0.185–0.187; SSIM drops when shaping or building-aware partitioning is removed (0.828 →\rightarrow 0.825 / 0.824). We add the same ablation on SMBU, together with a Base that disables all three components on an otherwise identical pipeline:

Variant (SMBU)PSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
Full24.9020.8090.171
w/o building-aware24.8620.8070.174
w/o planarity24.9020.8080.173
w/o shaping24.8350.8070.173
Base (all removed)24.8020.7960.181

On both scenes, every single removal raises LPIPS. Removing all three components together lowers every metric (PSNR 24.902 →\rightarrow 24.802, SSIM 0.809 →\rightarrow 0.796, LPIPS 0.171 →\rightarrow 0.181, a 5.8% relative increase in LPIPS). The joint effect is clearest in SSIM and LPIPS, where it is several times larger than that of any single removal (at most 0.002 SSIM and 0.003 LPIPS). This suggests that the components overlap: Gaussian shaping and planarity consistency both push Gaussians toward well-oriented, surface-aligned shapes, so either one recovers part of the effect when the other is removed. Beyond end-task metrics, the measurements in Q3 verify the partitioning mechanism directly.

Q3 – Mechanism measurements (SMBU, 103 buildings).

MetricInitial partition Π0\Pi^{0}Refined Π∗\Pi^{\ast}
Buildings crossing a block boundary181
Block load (max / mean primitives)1.0171.101

Building-aware refinement eliminates 17 of the 18 cross-block buildings while block loads remain balanced. With δ=0.3\delta = 0.3 bounding each reassignment, the largest cumulative block change is 10.5%. The one remaining case is a large vegetated hillside region; merging it would change a block's load by 65%, so the budget rejects it as intended. This count is also the relevant boundary measure: seams in block-wise reconstruction arise where an independently optimized block boundary cuts through a structure, and since each Gaussian is owned by exactly one block after merging, it gives where such a seam can fall on a building. For planar alignment, the Gaussians contributing to each detected plane instance in the trained SMBU model are coherently oriented: over 100 views and 869 instances, the per-view median resultant length of their weighted normals is 0.88–0.94, and no instance falls below 0.5.

W5 / Q4 – Cost (SMBU, 500 training images; modified RTX 4080 SUPER GPUs with 32 GB VRAM each; four GPUs for parallel reconstruction, one block per GPU):

ItemCost
ZeroPlane plane priors5.2 min for 500 images
Depth-Anything-V2 depth priors5.5 min for 500 images
Building extraction + partition + view assignment2.0 min
Plane-instance association + loss252 →\rightarrow 423 ms/iter at 10.4M Gaussians (largest block), active from step 6k; peak allocated GPU memory 17.9 →\rightarrow 20.8 GB
Total reconstruction, including all preprocessing≈9\approx 9 h on 4 GPUs (one block per GPU)
Final model13.2M Gaussians, 23.1 FPS at 1600×10641600\times1064

Together, the two prior-inference stages take 10.7 minutes, about 2% of the 9-hour total. Building extraction, partitioning and view assignment add another 2 minutes. The main overhead comes from plane-instance association and planarity regularization, which increase per-iteration time from 252 to 423 ms at the reported model size. All priors are computed once per scene and are not needed at rendering time.

Q1 – Geometric evaluation. Our claim concerns novel-view synthesis quality in large urban scenes, and NVS metrics on held-out views directly test the geometric changes we make. Every surface is scored from camera poses not used in training; a primitive with a wrong position or orientation can still reproduce the training images but appears smeared or displaced from a new pose, so geometric errors on facades show up as novel-view error, most visibly in the structure-sensitive SSIM and LPIPS. Consistent with this, removing planarity consistency or Gaussian shaping raises LPIPS on both Residence and SMBU. We keep finite-thickness primitives to preserve appearance on non-planar content such as vegetation and terrain. The final version will state this scope explicitly in the abstract and introduction.

We hope these additions address the reviewer's concerns, and we would be glad to clarify any remaining points.

Reviewer_KDoYOpenReview ↗

Mismatch between review and response posted

Could you kindly make sure that your intended response was posted under the correct review and directly address my concern?

Reviewer_Js3tOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UrbanBP-GS combines building-aware scene partitioning with planarity-constrained 3D Gaussian optimization for large-scale urban novel-view synthesis. It refines primitive ownership using connected building regions and regularizes Gaussian orientations using detected planar regions. Evaluation covers four urban scenes, with remove-one ablations on Residence.

Critical strengths
  1. The paper addresses a relevant practical problem. Preserving building continuity while controlling block sizes is a useful consideration for independently optimized urban reconstructions.
  2. The combination of structure-aware partitioning and planar regularization is intuitively reasonable. Retaining a 3D Gaussian representation while exploiting planar regions could help accommodate both regular building surfaces and irregular scene content.
Critical weaknesses
  1. Writing and presentation lack sufficient care. The logical organization is difficult to follow in places. For example, the paragraph beginning at line 065 discusses partitioning and building connectivity, whereas the paragraph beginning at line 071 abruptly switches to primitive dimensionality; the relationship between these arguments is insufficiently developed. Eqs. (3) and (4) each stack several definitions across multiple lines under a single equation number, making the construction harder to follow and audit. Small annotations in Figure 2 and low-contrast changes in Figure 3 further hinder readability. Together with repetitive, formulaic descriptions, these issues make the manuscript appear insufficiently edited and weaken the clarity of its technical argument.

  2. The validity of the motivation is insufficiently established. The paper treats building fragmentation and the limited flexibility of 2D primitives as key bottlenecks, but does not convincingly demonstrate that these are the causes of the observed rendering errors. Preserving building ownership does not by itself establish better reconstruction, and the benefit of retaining finite-thickness primitives needs a controlled comparison with block-wise 2DGS and standard 3DGS under matched supervision. Moreover, the algorithmic advances over PG-SAG [1] and PGSR-style regularization [2] remain unclear; a difference in target tasks alone does not establish novelty.

  3. The experimental setup and comparison choices are insufficiently justified. The evaluation does not adequately connect the geometry-oriented motivation to the novel-view-synthesis results. If the authors intend to demonstrate improved geometric reconstruction, CityGaussianV2 alone does not adequately represent subsequent developments: CityGS-X (ICCV 2025) [3], UrbanGS (ICLR 2026) [4], and MetroGS (CVPR 2026) [5] should be discussed, with appropriate comparisons and geometric metrics such as accuracy, completeness, and F-score. If the main contribution is novel-view synthesis, the discussion and evaluation still omit Momentum-GS (ICCV 2025) [6], MixGS (NeurIPS 2025) [7], and SigGS (ICLR 2026) [8]. The paper needs a clearly justified selection of relevant baselines and matched protocols to establish its position in the literature.

  4. The experimental evidence is not sufficiently convincing. Although Table 1 reports improvements over the selected baselines, the evidence is insufficient to attribute these gains to the proposed mechanisms or establish their practical significance. On Residence, building-aware refinement and planarity consistency improve PSNR by 0.165 and 0.127 dB, with LPIPS gains of 0.003 and 0.001, respectively. Ablations cover only one scene, lack repeated-run variability, and omit a common-backbone baseline with both proposed components disabled.

  5. The cost of additional prior inference is not evaluated. The method requires additional prior inference, including ZeroPlane plane detection [9] and monocular depth estimation. However, the paper gives no breakdown of the time or memory required to generate these priors, and reports neither total reconstruction time including preprocessing nor the overhead of plane-instance association during optimization. Without this accounting, it is unclear whether the reported quality gains justify the additional inference and processing costs, or whether comparisons use comparable computational budgets.

Ref: [1] T. Wang, X et al., "PG-SAG: Parallel Gaussian Splatting for Fine-Grained Large-Scale Urban Buildings Reconstruction via Semantic-Aware Grouping," arXiv preprint arXiv:2501.01677, 2025. [2] D. Chen et al., "PGSR: Planar-Based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction,"  IEEE Transactions on Visualization and Computer Graphics 31.9 (2024): 6100-6111. [3] Y. Gao et al., "CityGS-X: A Scalable Architecture for Efficient and Geometrically Accurate Large-Scale Scene Reconstruction," in Proc. IEEE/CVF Int. Conf. Computer Vision (ICCV), 2025. [4] C. Li et al., "UrbanGS: A Scalable and Efficient Architecture for Geometrically Accurate Large-Scene Reconstruction," in Int. Conf. Learning Representations (ICLR), 2026. [5] K. Chen et al., "MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale Scenes," in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2026. [6] J. Fan, W. Li, Y. Han, T. Dai, and Y. Tang, "Momentum-GS: Momentum Gaussian Self-Distillation for High-Quality Large Scene Reconstruction," in Proc. IEEE/CVF Int. Conf. Computer Vision (ICCV), 2025, pp. 25250-25260. [7] C. Liu, H. Wang, L. Yu, and G.-S. Xia, "Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting," in Advances in Neural Information Processing Systems (NeurIPS), vol. 38, 2025. [8] W. Xue et al., "Signal Structure-Aware Gaussian Splatting for Large-Scale Scene Reconstruction," in Int. Conf. Learning Representations (ICLR), 2026. [9] J. Liu, R. Yu, S. Chen, S. X. Huang, and H. Guo, "Towards In-the-Wild 3D Plane Reconstruction from a Single Image," in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2025.

Questions for authors
  1. This question arises from the motivation around 2DGS-based large-scale reconstruction, where geometric accuracy is typically a central evaluation criterion. Why are the experiments limited to NVS metrics? Can the authors provide geometric evaluations, or explain why NVS metrics alone adequately validate the proposed changes to primitive representation and planar constraints?
  2. Can matched-backbone ablations with neither module, each module individually, and both modules be provided, together with repeated-run variability on representative scenes?
  3. Can the authors directly quantify fragmentation, boundary consistency, planar alignment, and final block imbalance to verify the claimed mechanisms?
  4. What are the time and memory costs of prior inference and plane-instance association, and the total reconstruction cost including all preprocessing?
2 comments

Response to Reviewer (1/2)

We thank the reviewer for the detailed feedback, and for recognizing that preserving building continuity while controlling block sizes is a useful consideration and that combining structure-aware partitioning with planar regularization is intuitively reasonable. We respond below with new measurements and will incorporate all of them in the final version.

W1 – Presentation. We will restructure the introduction around a single question: which structures must remain intact within an independently optimized block, and which primitive represents them faithfully. The partitioning argument (L065) will then lead directly into the primitive argument (L071). Eqs. (3)–(4) will be split into separately numbered definitions, Fig. 2 annotations enlarged, Fig. 3 replaced by a high-contrast difference map, and repetitive descriptions condensed.

W2 – Algorithmic differences from PG-SAG and PGSR.

(i) Partitioning vs. PG-SAG. PG-SAG groups the scene into building-level sub-regions using semantic masks from a cross-modal segmentation model, then reconstructs each group in parallel. Our refinement needs no semantic model. It operates directly on primitive ownership in BEV: connected building regions are extracted from the coarse geometry, and each region is merged into the block it overlaps most, only if every affected block's relative primitive-count change stays within a budget δ\delta. This makes the fragmentation–load trade-off explicit while constraining each reassignment. On SMBU, buildings crossing a block boundary drop from 18 to 1, while the largest cumulative block change is 10.5% (Q3).

(ii) Planar regularization vs. PGSR. PGSR promotes planar Gaussians by minimizing their smallest scale and enforces single-view depth–normal and multi-view geometric consistency for surface reconstruction. We instead (a) keep finite-thickness 3D Gaussians: the shaping term stops once a primitive reaches a bounded aspect ratio (ρ≤τs\rho \le \tau_s) instead of driving its smallest scale to zero, so vegetation and terrain retain non-zero thickness; and (b) impose orientation consistency only among Gaussians that actually contribute to pixels of the same detected plane instance, identified through per-pixel contributor IDs from an extended rasterizer. Regularization therefore follows instance-level planar structure rather than applying uniformly to all primitives.

(iii) Finite thickness vs. 2DGS. We keep finite-thickness Gaussians because published results show that the 2DGS route mainly benefits geometry rather than rendering. 2DGS substantially improves surface reconstruction while reporting appearance quality competitive with 3DGS. CityGaussianV2, the 2DGS-based successor of CityGaussian, shows the same trade-off in its own evaluation, where both methods are evaluated under one protocol: on GauU-Scene it reports higher geometric F1 than CityGaussian (0.501 vs. 0.449) but lower PSNR and SSIM and higher LPIPS on all three scenes (e.g., Russian Building: 24.12 / 0.784 / 0.196 vs. 24.37 / 0.808 / 0.163). Since urban scenes also contain substantial non-planar content such as vegetation and terrain, we retain the third scale for novel-view synthesis and add planarity only where planes are detected.

W3 – Baselines. We will add Momentum-GS, MixGS, SigGS, CityGS-X, UrbanGS and MetroGS to Table 1. Numbers are taken from each paper (PSNR / SSIM / LPIPS); protocol differences, such as MetroGS's per-image appearance embeddings, will be listed in the caption.

MethodBuildingRubbleResidenceMatrixCity-Aerial
Momentum-GS23.65 / 0.813 / 0.19426.66 / 0.826 / 0.20023.37 / 0.828 / 0.19629.11 / 0.881 / 0.180
MixGS23.03 / 0.771 / 0.26126.66 / 0.792 / 0.26723.39 / 0.815 / 0.219–
SigGS (SIG-L)22.86 / 0.778 / 0.22527.35 / 0.843 / 0.18923.46 / 0.840 / 0.19429.04 / 0.888 / 0.184
CityGS-X22.76 / 0.817 / 0.19126.15 / 0.823 / 0.21022.44 / 0.819 / 0.19427.58 / – / –
UrbanGS22.82 / 0.802 / 0.20826.25 / 0.791 / 0.21022.48 / 0.823 / 0.205–
MetroGS23.06 / 0.787 / 0.17327.48 / 0.826 / 0.14723.38 / 0.824 / 0.16627.52 / 0.854 / 0.167
Ours22.63 / 0.787 / 0.19526.09 / 0.803 / 0.20323.32 / 0.828 / 0.18427.44 / 0.871 / 0.139

UrbanBP-GS achieves the lowest LPIPS on MatrixCity-Aerial (0.139 vs. 0.167–0.184) and lower LPIPS than MixGS and UrbanGS on every scene where they report results. Our contributions act on how the scene is partitioned and how primitives are oriented, and are complementary to the representation and training designs of these methods (self-distillation, holistic decoding, signal-aware scheduling, appearance-decoupled 2DGS). PG-SAG and PGSR are discussed in W2 and will be added to Related Work.

Response to Reviewer (2/2)

W4 / Q2 – Ablations. Table 2 removes each component from the full model on Residence while keeping the coarse model, training schedule and evaluation identical. Removing any component lowers PSNR and raises LPIPS: Gaussian shaping costs 0.284 dB, building-aware partitioning 0.165 dB and planarity consistency 0.127 dB, with LPIPS rising from 0.184 to 0.185–0.187; SSIM drops when shaping or building-aware partitioning is removed (0.828 →\rightarrow 0.825 / 0.824). We add the same ablation on SMBU, together with a Base that disables all three components on an otherwise identical pipeline:

Variant (SMBU)PSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
Full24.9020.8090.171
w/o building-aware24.8620.8070.174
w/o planarity24.9020.8080.173
w/o shaping24.8350.8070.173
Base (all removed)24.8020.7960.181

On both scenes, every single removal raises LPIPS. Removing all three components together lowers every metric (PSNR 24.902 →\rightarrow 24.802, SSIM 0.809 →\rightarrow 0.796, LPIPS 0.171 →\rightarrow 0.181, a 5.8% relative increase in LPIPS). The joint effect is clearest in SSIM and LPIPS, where it is several times larger than that of any single removal (at most 0.002 SSIM and 0.003 LPIPS). This suggests that the components overlap: Gaussian shaping and planarity consistency both push Gaussians toward well-oriented, surface-aligned shapes, so either one recovers part of the effect when the other is removed. Beyond end-task metrics, the measurements in Q3 verify the partitioning mechanism directly.

Q3 – Mechanism measurements (SMBU, 103 buildings).

MetricInitial partition Π0\Pi^{0}Refined Π∗\Pi^{\ast}
Buildings crossing a block boundary181
Block load (max / mean primitives)1.0171.101

Building-aware refinement eliminates 17 of the 18 cross-block buildings while block loads remain balanced. With δ=0.3\delta = 0.3 bounding each reassignment, the largest cumulative block change is 10.5%. The one remaining case is a large vegetated hillside region; merging it would change a block's load by 65%, so the budget rejects it as intended. This count is also the relevant boundary measure: seams in block-wise reconstruction arise where an independently optimized block boundary cuts through a structure, and since each Gaussian is owned by exactly one block after merging, it gives where such a seam can fall on a building. For planar alignment, the Gaussians contributing to each detected plane instance in the trained SMBU model are coherently oriented: over 100 views and 869 instances, the per-view median resultant length of their weighted normals is 0.88–0.94, and no instance falls below 0.5.

W5 / Q4 – Cost (SMBU, 500 training images; modified RTX 4080 SUPER GPUs with 32 GB VRAM each; four GPUs for parallel reconstruction, one block per GPU):

ItemCost
ZeroPlane plane priors5.2 min for 500 images
Depth-Anything-V2 depth priors5.5 min for 500 images
Building extraction + partition + view assignment2.0 min
Plane-instance association + loss252 →\rightarrow 423 ms/iter at 10.4M Gaussians (largest block), active from step 6k; peak allocated GPU memory 17.9 →\rightarrow 20.8 GB
Total reconstruction, including all preprocessing≈9\approx 9 h on 4 GPUs (one block per GPU)
Final model13.2M Gaussians, 23.1 FPS at 1600×10641600\times1064

Together, the two prior-inference stages take 10.7 minutes, about 2% of the 9-hour total. Building extraction, partitioning and view assignment add another 2 minutes. The main overhead comes from plane-instance association and planarity regularization, which increase per-iteration time from 252 to 423 ms at the reported model size. All priors are computed once per scene and are not needed at rendering time.

Q1 – Geometric evaluation. Our claim concerns novel-view synthesis quality in large urban scenes, and NVS metrics on held-out views directly test the geometric changes we make. Every surface is scored from camera poses not used in training; a primitive with a wrong position or orientation can still reproduce the training images but appears smeared or displaced from a new pose, so geometric errors on facades show up as novel-view error, most visibly in the structure-sensitive SSIM and LPIPS. Consistent with this, removing planarity consistency or Gaussian shaping raises LPIPS on both Residence and SMBU. We keep finite-thickness primitives to preserve appearance on non-planar content such as vegetation and terrain. The final version will state this scope explicitly in the abstract and introduction.

We hope these additions address the reviewer's concerns, and we would be glad to clarify any remaining points.

Reviewer_37fPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents UrbanBP-GS, a method designed to improve large-scale urban view synthesis using block-partitioned 3D Gaussian Splatting. The framework relies on two primary mechanics. First, building-aware partitioning adjusts scene blocks using connected building structures extracted from coarse geometry, which helps prevent single buildings from being split across boundaries while managing shifts in total primitive counts. Second, a planarity constraint aligns the orientations of 3D Gaussians associated with detected planar surfaces without discarding their three optimizable scale parameters, allowing the model to represent both flat and complex geometry.

The model is benchmarked across four large urban scenes against several NeRF and Gaussian baselines. UrbanBP-GS delivers the lowest LPIPS across every test scene and leads in PSNR on three of them. Qualitative outputs and ablation tests are provided using the Residence dataset.

Critical strengths
  1. The paper tackles a clear issue in block-based urban modeling, where training connected buildings in separate partitions creates noticeable artifacts across block edges. Using structural geometry to group building regions is a practical fix, and deriving these shapes from coarse representations avoids relying on manual annotations.

  2. The performance remains strong across standard test cases: the technique scores the best LPIPS across all four datasets, alongside clear PSNR gains on the Building, Rubble, and Residence scenes. The visual figures show cleaner facade boundaries and better recovery of thin structures.

Critical weaknesses
  1. The math of the planarity constraint seems flawed. In Equation (4), the authors claim the alignment function is sign-invariant. However, the shared planar normal is computed using a weighted average of signed Gaussian vectors. If two vectors within the same plane point in opposite directions (n\mathbf{n} and −n-\mathbf{n}), their weighted average cancels out to zero. While the squared dot product applied later handles sign changes for a fixed reference vector, the full equation remains sensitive to individual sign flips. This creates a real issue because a Gaussian covariance matrix does not enforce a single orientation for its local axes. Can the authors please clarify this point?

  2. Following my previous point, Equation (2) contains a similar issue. The relative primitive variation is signed, yet the acceptance check caps the upper bound without using absolute values. As a result, the rule allows large drops in primitive counts that exceed the intended budget limit. Furthermore, the threshold rule meant to shape the Gaussians compares only the longest and shortest scales. This fails to ensure that the shortest axis is unique, as two matching small scale values can satisfy the check simultaneously.

  3. The experiments do not clearly prove how much each individual component helps. The ablation study covers only the Residence scene. Here, removing planarity alignment lowers PSNR from 23.315 to 23.188, keeps SSIM fixed at 0.828, and shifts LPIPS from 0.184 to 0.185. Removing building-aware refinement drops PSNR to 23.150 and moves LPIPS to 0.187. These changes are small compared to the gaps seen between the full system and the external baselines.

  4. The experiments lack a controlled baseline with both shaping and planarity losses disabled, making it hard to separate the gains of the new updating scheme from standard 3D Gaussian Splatting. Additionally, the building-aware partitioning is missing direct data on building fragmentation, block imbalance, total training times, and peak memory usage. The claims regarding improved planar and detailed geometry are supported only by whole-image rendering metrics rather than structural or regional evaluations.

  5. The related work section skips key papers on large-scale Gaussian reconstruction. Specifically, reference [1] handles visibility-based view selection and workload rebalancing, while reference [2] combines coarse Gaussian setups with object and image partitioning. Discussing these works is necessary to clarify how the proposed partitioning scales and handles computational loads.

References

[1] Robust and Efficient 3D Gaussian Splatting for Urban Scene Reconstruction

[2] Gaussian Splatting for Large-Scale Aerial Scene Reconstruction From Ultra-High-Resolution Images

Questions for authors
  1. How does the algorithm resolve sign ambiguity when computing the reference planar normal? Specifically, what happens when Gaussians on the same plane have opposite local direction vectors?

  2. Is the use of signed variation in Equation (2) intentional? If so, how are large reductions in primitive counts managed, and what is the final balance across blocks?

  3. Can you show a baseline 3DGS comparison with shaping and planarity constraints disabled, alongside geometric ablation tests on a second scene?

  4. How does this partitioning pipeline differ in operation from the scene and view partitioning techniques in references [1] and [2] ? Is it possible to compare these methods ?

3 comments

Response to Reviewer (1/2) – Mathematical Clarifications and Partition Balance

We thank the reviewer for the careful evaluation and for recognizing the practical value of building-aware partitioning and the rendering improvements of UrbanBP-GS. We address the concerns below with additional analyses of our existing results and clarify the mathematical formulation, partition balance, ablation protocol, and relationship to prior work.

W1 / Q1 – Sign ambiguity of the reference normal (Eq. 4).

We agree that squaring the inner product alone does not guarantee sign invariance when the reference direction is constructed by averaging signed vectors. We therefore examined whether this ambiguity causes cancellation in the trained model. On SMBU, we analyzed 100 training views, 869 plane instances, and 89.6M Gaussian–plane memberships:

MeasurementValue
Resultant length of weighted signed mean (per-view medians)0.88–0.94
Instances with resultant length below 0.50%
Angle to the unoriented principal axis (per-view medians)0.05°–0.29°
Instances with angle above 10°0%

These measurements show that severe cancellation does not occur in the evaluated model: the signed mean remains close to the corresponding unoriented principal direction for all measured instances. Nevertheless, we acknowledge that this is empirical stability rather than a mathematical guarantee.

To clarify how strict sign invariance can be enforced, we provide an unoriented tensor formulation:

Ni=∑maimrimrim⊤N_i = \sum_m a_{im} r_{im} r_{im}^{\top}

Tvp=∑i∈GvpωiNi∑i∈GvpωiT_{vp} = \frac{\sum_{i \in G_{vp}} \omega_i N_i}{\sum_{i \in G_{vp}} \omega_i}

Lplanaritytensor=mean⁡v,p(1−λmax⁡(Tvp))L_{\mathrm{planarity}}^{\mathrm{tensor}} = \operatorname{mean}_{v,p}\left(1-\lambda_{\max}(T_{vp})\right)

Unlike signed-vector aggregation, this formulation depends on the local axes through their outer products and is therefore invariant to arbitrary axis sign flips. On the same SMBU memberships, random covariance-preserving axis flips change this tensor loss by at most 3.7×10−153.7 \times 10^{-15} in double precision. This verifies the numerical invariance of the alternative formulation, rather than the optimization behavior of the original signed-vector loss. We will clarify this distinction and include the invariant formulation in the appendix.

W2 / Q2 – Primitive-count variation and shortest-axis ambiguity.

We agree that an upper bound on a signed primitive-count change would not prevent large reductions. The intended variation measure, which is also used in our implementation, is the absolute relative change of each affected block:

Δk(j)=∣N~k(j)−Nk(j−1)∣max⁡(Nk(j−1),1)\Delta_k^{(j)} = \frac{|\widetilde{N}_k^{(j)}-N_k^{(j-1)}|}{\max(N_k^{(j-1)},1)}

max⁡kΔk(j)≤δ\max_k \Delta_k^{(j)} \leq \delta

The implementation checks both receiving and donor blocks, thereby constraining decreases as well as increases. We will correct the notation in Eq. (2) to make this explicit.

Importantly, the budget constrains individual refinement operations, not cumulative deviation. We therefore measured the final distribution on SMBU (4 blocks, 103 extracted candidate regions, δ=0.3\delta=0.3):

MeasurementInitial partitionRefined partitionAfter training
Cross-block regions181–
Per-block primitives (M)2.28 / 2.31 / 2.27 / 2.222.50 / 2.37 / 2.03 / 2.173.64 / 3.77 / 2.81 / 3.00
Max/mean load1.0171.1011.142
Min/mean load0.9770.8960.850

The largest cumulative block-count change is 10.5%, and fragmented extracted-region area decreases from 45.7% to 27.3%. The remaining cross-block region corresponds to a large vegetated hillside included in the extracted candidates. Consolidating it would change a block's load by approximately 65% and increase the maximum load to 1.60 times the mean. The budget therefore prevents this disproportionately expensive reassignment. After densification to 13.2M Gaussians, the maximum load remains at 1.142 times the mean, indicating that the refined partition maintains reasonably balanced workloads through training.

Shortest-axis ties. We agree that the scale-ratio criterion alone does not guarantee a unique shortest axis. The softmax in Eq. (3) makes the axis-selection weights continuous with respect to scale changes, but does not by itself eliminate axis-sign ambiguity.

The tensor formulation provides a representation-independent alternative. When s1=s2s_1=s_2, the two tied axes receive equal weights, giving

Ni=a(r1r1⊤+r2r2⊤)+a3r3r3⊤N_i = a(r_1r_1^{\top}+r_2r_2^{\top})+a_3r_3r_3^{\top}

This tensor is invariant to rotations within the tied eigenspace and to arbitrary axis sign flips. Thus, although a unique shortest-axis direction is not geometrically identifiable at an exact tie, the tensor and its scalar regularization objective remain well-defined without arbitrarily selecting a normal direction. We will clarify this distinction in the revised manuscript.

Response to Reviewer (2/3) – Ablation Evidence and Computational Cost

W3 / Q3 – Component contributions and common-backbone ablation.

We agree that the Residence-only ablation provides limited evidence of cross-scene consistency. We therefore additionally evaluated the components on SMBU, including a common-backbone Base with all three proposed components disabled. All variants share the same coarse checkpoint, data split, monocular inverse-depth supervision, number of blocks, and 60k training steps per block.

Variant (SMBU)PSNR ↑SSIM ↑LPIPS ↓
Full24.9020.8090.171
w/o building-aware24.8620.8070.174
w/o planarity24.9020.8080.173
w/o shaping24.8350.8070.173
Base (all removed)24.8020.7960.181

The Base disables both geometric regularizers and building-aware refinement, providing a controlled reference for their combined contribution within the same training backbone.

Removing all three components decreases PSNR by 0.100 dB and SSIM by 0.013, while increasing LPIPS by 0.010. Every individual removal also increases LPIPS on both Residence and SMBU. In particular, removing planarity consistency produces a modest but consistent LPIPS degradation, although its PSNR effect is negligible at the reported precision on SMBU.

The joint degradation in SSIM and LPIPS exceeds the sum of the individual remove-one degradations, suggesting complementary or non-additive interactions among the components. The common-backbone comparison supports the contribution of the proposed framework under matched initialization, supervision, and training iterations, without conflating it with differences in these settings.

W4 – Computational cost and structural evidence.

We additionally measured the pipeline on SMBU (500 training images, 1600x1064), using four RTX 4080 SUPER GPUs modified to 32 GB memory each.

ItemMeasured cost
ZeroPlane plane-mask prediction5.2 min
Depth-Anything-V2 (ViT-L) inverse depth5.5 min
Coarse training (30k steps)1.37 h
Building extraction, partitioning, and view assignment2.0 min
Iteration time, planarity off to on252 to 423 ms
Peak allocated GPU memory, off to on17.9 to 20.8 GB
End-to-end reconstructionApproximately 9 h on 4 GPUs
Final reconstruction13.2M Gaussians, 23.1 FPS at 1600x1064

The iteration-time measurements are taken on the largest block model (10.4M Gaussians before ownership filtering), with contributor tracking and planarity supervision active from step 6k. Their activation increases iteration time by approximately 68% and peak allocated memory by approximately 16%.

Plane-mask prediction, depth estimation, and geometry processing together take approximately 13 minutes, less than 3% of the end-to-end reconstruction time. We acknowledge the additional training-time cost of contributor tracking and plane regularization; however, these operations are not required during rendering.

Structural evidence. Beyond whole-image metrics, the measurements in W2 directly quantify the partitioning mechanism: cross-block extracted regions decrease from 18 to 1, while the maximum refined-block load remains at 1.101 times the mean.

The normal-resultant measurements in W1 characterize orientation coherence among Gaussians associated with detected plane instances. They do not independently establish ground-truth surface-normal accuracy.

We distinguish these measurements from rendering quality: the partition statistics directly support structural consolidation, the resultant statistics characterize orientation consistency, and the held-out-view SSIM and LPIPS results provide complementary evidence of rendering improvements. In particular, removing either planarity consistency or Gaussian shaping increases LPIPS on both Residence and SMBU. We will clarify this distinction rather than treating whole-image rendering metrics alone as direct proof of geometric accuracy.

Response to Reviewer (3/3) – Relationship to Prior Work

W5 / Q4 – Relationship to [1] and [2].

We thank the reviewer for highlighting these closely related works, which we will include in Related Work. We agree that scene partitioning and visibility-based view selection are established techniques. Our contribution lies in the specific structural refinement objective and its integration with local geometric regularization.

[1] Yuan et al. (ICCV 2025) employ scene partitioning, visibility-aware image assignment, and partition merging or subdivision to balance training-image counts, together with a controllable LOD mechanism for regulating Gaussian density. UrbanBP-GS likewise considers camera location and visibility for view assignment, but differs in its refinement criterion: instead of balancing image counts, it adjusts primitive ownership to consolidate connected building structures while constraining per-update primitive-count variation. Thus, the two methods address different aspects of large-scale reconstruction. Their strategies could also be complementary, since visibility-based selection and LOD control can operate within structurally refined blocks.

[2] Sun et al. (CGF 2025) combine object-space Gaussian partitioning with image-space subdivision and depth-based visibility masks, enabling ultra-high-resolution aerial reconstruction under constrained GPU memory. Their primary objective is to limit the memory footprint of individual subtasks. Our building-aware refinement instead targets structural fragmentation across block boundaries while maintaining primitive-count balance, without relying on an equivalent high-resolution image-space subdivision mechanism.

We will clarify these differences in terms of partitioning objectives, balancing criteria, and computational constraints. Although the methods address related large-scale reconstruction problems, their published training configurations, resource budgets, and evaluation protocols are not fully matched. We therefore avoid interpreting cross-paper metrics as a controlled comparison of partitioning strategies.

Instead, our common-backbone ablations establish the contribution of the proposed components under matched settings, while the direct fragmentation and load measurements quantify the intended effect of building-aware refinement.

We thank the reviewer again for the constructive suggestions. We believe these additional analyses and clarifications more precisely establish the contributions, empirical behavior, and practical tradeoffs of UrbanBP-GS.

Reviewer_WC5SOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UrbanBP-GS for large-scale urban novel-view synthesis. The method combines primitive-count-based initialization with building-aware ownership refinement, then optimizes each block using Gaussian shaping and plane-instance-based orientation consistency. Independently trained blocks are merged using the refined ownership map. Experiments on four urban scenes report competitive rendering quality. The problem is relevant, but the orientation-alignment formulation and the evidence supporting the proposed partitioning and geometric mechanisms need substantial clarification.

Critical strengths

The paper addresses practical problems in urban Gaussian splatting, including buildings split across blocks and inconsistent planar appearance. Combining geometry-aware ownership with local orientation regularization is a coherent approach to these problems. The evaluation includes several relevant large-scale baselines and four scenes, with the highest reported PSNR on three scenes and the lowest LPIPS on all four. The manuscript is generally well organized, and the implementation schedule and remove-one ablations provide useful information. The authors also explicitly acknowledge that the refinement budget limits individual reassignments rather than cumulative load deviation.

Critical weaknesses
  1. Equation (4) does not establish the claimed sign invariance of the planarity objective. For a fixed reference direction, the squared inner product is sign-invariant. The reference itself, however, is computed by averaging signed vectors. Opposite normals with equal weights can cancel even when they represent the same plane, leaving the shared direction unstable or undefined. This is a substantive ambiguity because reversing a plane normal should not change the underlying geometry. The authors should specify a consistent orientation convention or an unoriented aggregation method and test whether equivalent normal-sign assignments produce equivalent optimization behavior.
  2. The shaping criterion in Section 3.3 does not guarantee a unique shortest axis. Scales of (0.1, 0.1, 1.0) have a small minimum-to-maximum ratio, yet the two shortest axes are tied. The soft axis combination in Eq. (3) may also depend on arbitrary axis signs that leave the Gaussian covariance unchanged. The authors should justify the direction estimator geometrically and test its behavior near scale ties.
  3. The partitioning results do not establish the tradeoff between building integrity and workload balance. Budgets on individual reassignments allow cumulative imbalance, and densification can change the loads further. The authors should report building fragmentation and per-block primitive distributions before refinement, after refinement, and after training. They should also show how these measurements vary with the refinement budget.
  4. The ablation study covers only Residence. The reported gains are approximately 0.165 dB for refinement and 0.127 dB for planarity consistency. Repeated runs, additional scenes, and measurements focused on block boundaries are needed to determine whether these modest gains are consistent.
  5. The plane-to-Gaussian association may impose incorrect constraints. Section 3.3 collects rendering contributors within each plane mask without accumulated-alpha ranking or cross-view plane matching. Weak background contributions and inconsistent masks could therefore assign a Gaussian to conflicting groups. The authors should specify the filtering thresholds and explain how they handle conflicting memberships. An association-error analysis and a comparison with an alpha-aware or confidence-weighted alternative would help determine whether the orientation regularizer remains reliable with imperfect masks.
  6. The baseline protocol does not sufficiently isolate the proposed contributions. It is unclear which results were reproduced using identical image splits, depth supervision, training budgets, and block configurations. A common training pipeline with matched priors and resources is needed to determine how much of the gain comes from the proposed partitioning and regularization. A baseline that removes the proposed components together would help assess their combined effect. A controlled comparison with planar primitives would also help substantiate the claimed advantage of retaining three optimizable scales for irregular geometry.
  7. The paper does not quantify the costs of coarse training, geometry processing, plane-mask prediction, or contributor tracking. This makes practical scalability and the value of the rendering gains difficult to assess. The authors should report preprocessing and training time, peak GPU memory, final primitive counts, rendering speed, and hardware, along with the complete depth-estimation and loss settings.
Questions for authors

How does the shared-direction calculation avoid cancellation between equivalent opposite normals? How is a stable direction obtained when the two smallest Gaussian scales coincide? Can the authors show that building-aware refinement improves building integrity without unacceptable cumulative load imbalance, including after densification? Do the rendering gains hold across scenes and repeated runs with matched supervision and computational budgets? How are conflicting plane memberships handled, and how much additional preprocessing time, training cost, and GPU memory does the full pipeline require?

4 comments
AuthorsOpenReview ↗

Response to Reviewer (1/4) – Sign Invariance and Shortest-Axis Ambiguity

We thank the reviewer for the careful evaluation and for recognizing the practical relevance of building-aware refinement and orientation regularization. We address the concerns below with additional measurements and mathematical clarifications, which will be incorporated into the revised manuscript.

W1 / Q1 – Sign invariance of the shared direction (Eq. 4).

We agree that squaring the inner product does not guarantee sign invariance when the reference direction is obtained by averaging signed vectors. We therefore examined whether this ambiguity leads to cancellation in the trained model. On SMBU, we analyzed 100 training views, 869 plane instances, and 89.6M Gaussian–plane memberships:

MeasurementValue
Weighted signed-mean resultant length (per-view medians)0.88–0.94
Instances with resultant length below 0.50%
Angle to the unoriented principal axis (per-view medians)0.05°–0.29°
Instances with angle above 10°0%

These results show that severe cancellation does not occur in the evaluated model: none of the measured instances has a resultant length below 0.5 or a deviation exceeding 10° from the unoriented principal direction. However, empirical stability does not establish sign invariance for arbitrary axis assignments.

To make this distinction explicit, we provide an unoriented tensor formulation:

Ni=∑maimrimrim⊤N_i = \sum_m a_{im} r_{im} r_{im}^{\top}

Tvp=∑i∈GvpωiNi∑i∈GvpωiT_{vp} = \frac{\sum_{i \in G_{vp}} \omega_i N_i}{\sum_{i \in G_{vp}} \omega_i}

Lplanaritytensor=mean⁡v,p(1−λmax⁡(Tvp))L_{\mathrm{planarity}}^{\mathrm{tensor}} = \operatorname{mean}_{v,p} \left(1-\lambda_{\max}(T_{vp})\right)

Because each axis contributes through its outer product, the tensor loss is invariant to arbitrary local-axis sign flips. On the same SMBU memberships, random covariance-preserving axis flips change this loss by at most 3.7×10−153.7 \times 10^{-15} in double precision. This verifies the numerical invariance of the tensor formulation, rather than claiming that the original signed-vector formulation is invariant. We will clarify this distinction and include the invariant formulation in the appendix.

W2 / Q2 – Direction estimation near scale ties.

We agree that the scale-ratio criterion alone does not guarantee a unique shortest axis. The softmax selector in Eq. (3) makes the axis-selection weights continuous as scales approach a tie, but the resulting signed-vector direction can still depend on the axis representation.

The tensor formulation avoids this ambiguity. When s1=s2s_1=s_2, the two tied axes receive equal weights, giving

Ni=a(r1r1⊤+r2r2⊤)+a3r3r3⊤N_i = a(r_1 r_1^{\top}+r_2 r_2^{\top})+a_3 r_3 r_3^{\top}

This tensor is invariant to rotations within the tied eigenspace and to arbitrary axis sign flips. Although a unique shortest-axis direction is not geometrically identifiable at an exact tie, the tensor and its scalar regularization objective remain well-defined without selecting an arbitrary normal. We will clarify the limitations of the original direction estimator and the representation-independent property of the tensor formulation.

AuthorsOpenReview ↗

Response to Reviewer (2/4) – Building-Aware Partitioning and Ablation Studies

W3 / Q3 – Building integrity versus workload balance.

We measured the effect of refinement on SMBU (103 extracted candidate regions, 4 blocks) under different budgets δ\delta. Load is measured by the primitive count of each block relative to the mean:

BudgetCrossingsFrag. areaMax/meanMin/meanCumulative change
Initial partition1845.7%1.0170.977–
0.1–0.5 (paper: 0.3)127.3%1.1010.89610.5%
At least 0.6600.0%1.5960.48063.4%

(1) Structural integrity. All tested budgets in the interval [0.1,0.5][0.1,0.5], including our default δ=0.3\delta=0.3, produce the same refined partition. Refinement consolidates 17 of the 18 cross-block candidate regions. The remaining region corresponds to a large vegetated hillside included in the extracted candidates, rather than an ordinary building instance. It accounts for the remaining fragmented candidate-region area.

(2) Cumulative imbalance. The budget constrains individual reassignments, not cumulative load deviation. We therefore measured the latter explicitly. Per-block primitive counts change from 2.28 / 2.31 / 2.27 / 2.22 M to 2.50 / 2.37 / 2.03 / 2.17 M, with the largest cumulative change being 10.5%. The maximum load remains at 1.101 times the mean.

(3) Effect of the budget. Consolidating the remaining large region requires δ≥0.66\delta \geq 0.66 and substantially worsens balance: the maximum load reaches 1.596 times the mean, while the minimum falls to 0.480. The budget therefore prevents a disproportionately expensive reassignment rather than forcing every extracted region into a single block.

(4) After densification. The trained SMBU model contains 13.2M Gaussians after 60k steps. When grouped by the refined ownership map Π∗\Pi^*, the four blocks contain 3.64 / 3.77 / 2.81 / 3.00 M Gaussians, corresponding to growth factors of approximately 1.38–1.59. The maximum load ratio increases only from 1.101 to 1.142, and the minimum changes from 0.896 to 0.850. Thus, although the budget does not explicitly constrain post-training counts, the observed imbalance remains moderate after densification.

W4 / Q4 – Consistency and combined contribution of the components.

In the Residence remove-one ablation (Table 2), removing Gaussian shaping, building-aware refinement, and planarity consistency reduces PSNR by 0.284, 0.165, and 0.127 dB, respectively. LPIPS increases in all three cases (from 0.184 to 0.185–0.187).

We additionally evaluated the components on SMBU and included a common-backbone Base with all three proposed components disabled:

Variant (SMBU)PSNR ↑SSIM ↑LPIPS ↓
Full24.9020.8090.171
w/o building-aware24.8620.8070.174
w/o planarity24.9020.8080.173
w/o shaping24.8350.8070.173
Base (all removed)24.8020.7960.181

All variants share the same coarse checkpoint, data split, monocular inverse-depth supervision, number of blocks, and 60k steps per block.

Removing all three components decreases PSNR by 0.100 dB and SSIM by 0.013, while increasing LPIPS by 0.010. Every individual removal also increases LPIPS on both Residence and SMBU. The planarity term has no distinguishable PSNR effect at the reported precision on SMBU, but yields modest SSIM and LPIPS improvements.

The joint degradation in SSIM and LPIPS exceeds the sum of the individual remove-one degradations, suggesting complementary or non-additive interactions among the components. We do not interpret the smaller individual differences as statistically significant without repeated-run variance estimates.

The W3 measurements additionally verify the partitioning mechanism directly: refinement substantially reduces the number of extracted structural regions crossing block boundaries. This reduces potential sites of cross-block structural discontinuity, although the fragmentation statistics are not themselves measurements of boundary rendering error. Together, the two-scene ablations and direct partition measurements provide complementary evidence for the proposed design.

AuthorsOpenReview ↗

Response to Reviewer (3/4) – Plane–Gaussian Association and Baseline Protocol

W5 / Q5 – Plane–Gaussian association and conflicting memberships.

We clarify the filtering, association, and conflict-handling procedures.

(i) Visibility and opacity filtering. Contributor collection is performed within the rasterizer. A Gaussian is recorded for a pixel only when its per-pixel opacity satisfies α≥1/255\alpha \geq 1/255 and the ray transmittance has not fallen below 10−410^{-4}. These conditions exclude very low-opacity contributions and contributions behind heavily occluded surfaces. The procedure does not explicitly rank contributors by accumulated alpha-weighted visibility, and we will clarify this distinction.

(ii) Plane-instance filtering. For each view, we retain the 10 largest detected plane instances and discard masks smaller than 10 pixels. Contributor IDs are deduplicated and restricted to visible Gaussians. Instances associated with fewer than 3 Gaussians are discarded. These filters prevent very small masks and minimally supported instances from contributing to the regularization.

(iii) Conflicting memberships. Associations are constructed independently for each view, without persistent cross-view plane identities or hard global plane assignments. A Gaussian associated with multiple instances can therefore receive gradients from multiple instance-level terms. Each term is normalized by its total Gaussian weight, ∑iωi\sum_i \omega_i, so its scale does not grow directly with the number of associated Gaussians.

On the 100 sampled SMBU views, 78.6% of Gaussian–plane memberships correspond to Gaussians associated with only one plane instance within their view. The remaining memberships may participate in multiple instance-level constraints. We acknowledge that per-view normalization does not eliminate the effect of incorrect masks, but avoids committing a Gaussian to a persistent, potentially incorrect global plane identity.

The planarity weight also decays from approximately 0.148 at its activation step to 0.01 by the end of training, reducing its relative influence during later optimization. Together, visibility filtering, per-view association, and the decaying regularization weight limit the reliance on potentially imperfect plane assignments. We will explicitly document these implementation details without claiming that all association errors are eliminated.

W6 – Baseline protocol and planar primitives.

Baseline protocol. Table 1 reports baseline results quoted from the corresponding publications, rather than independently reproduced runs. We follow the reported evaluation resolution policy (4x downsampling on Mill19 and UrbanScene3D, and 1.2x on MatrixCity-Aerial) and the corresponding train–test splits.

However, published baselines may differ in their training supervision, optimization budgets, block configurations, and computational resources. We therefore distinguish cross-paper benchmark comparisons from controlled ablations and do not claim that all external baselines were trained under an identical pipeline.

For the latter, all UrbanBP-GS variants in Table 2 and the additional SMBU ablation share the same coarse checkpoint, data split, monocular inverse-depth supervision, block count, and 60k training steps per block. The SMBU Base removes building-aware refinement, Gaussian shaping, and planarity consistency together, providing a matched reference for their combined contribution within our training framework.

Planar primitives. Our motivation for retaining finite-thickness 3D Gaussians is to preserve representational flexibility in urban scenes containing both planar facades and irregular structures such as vegetation and terrain. Published comparisons indicate a tradeoff between surface reconstruction accuracy and rendering quality.

For example, CityGaussianV2, which adopts a 2DGS-based representation, reports higher geometric F1 than CityGaussian on GauU-Scene (0.501 versus 0.449), but lower PSNR and SSIM and higher LPIPS across its three evaluated scenes. On Russian Building, the reported PSNR/SSIM/LPIPS values are 24.12/0.784/0.196 for CityGaussianV2 and 24.37/0.808/0.163 for CityGaussian.

These published results motivate our choice to retain three optimizable Gaussian scales while applying planar constraints selectively. They are not a controlled comparison between planar and volumetric primitives within UrbanBP-GS, and we do not claim that our representation universally outperforms 2DGS in geometric accuracy. Our design instead targets novel-view rendering while accommodating both regular and irregular scene content.

AuthorsOpenReview ↗

Response to Reviewer (4/4) – Computational Cost and Implementation Details

W7 / Q5 – Computational cost and implementation settings.

We measured the pipeline on SMBU (500 training images, 1600x1064 resolution), using four RTX 4080 SUPER GPUs modified to 32 GB memory each, with PyTorch 2.0.1 and CUDA 11.8.

ItemMeasured cost
ZeroPlane plane-mask prediction5.2 min / 500 images
Depth-Anything-V2 (ViT-L) inverse depth5.5 min / 500 images
Coarse training (30k steps)1.37 h
Building extraction, partition, view assignment2.0 min
Iteration time, planarity off to on252 to 423 ms
Peak allocated GPU memory, off to on17.9 to 20.8 GB
Total reconstruction, including preprocessingApproximately 9 h on 4 GPUs
Final model13.2M Gaussians
Rendering speed at 1600x106423.1 FPS

The iteration-time measurements are taken on the largest block model (10.4M Gaussians before ownership filtering), with contributor tracking, plane association, and the planarity loss enabled from step 6k. Their activation increases the measured iteration time from 252 to 423 ms and peak allocated memory from 17.9 to 20.8 GB.

Plane-mask prediction, depth estimation, and geometry processing together require approximately 13 minutes, accounting for less than 3% of the approximately 9-hour end-to-end reconstruction. The four blocks are trained in parallel, one per GPU. We acknowledge the additional training-time overhead of contributor tracking and plane regularization; however, these operations are not required at rendering time.

Loss functions and schedules.

The photometric loss is

Lrgb=0.8L1+0.2(1−SSIM)L_{\mathrm{rgb}} = 0.8 L_1 + 0.2(1-\mathrm{SSIM})

The depth loss applies the same combination of L1 and SSIM terms to monocular inverse-depth supervision from Depth-Anything-V2 (ViT-L).

Let

h(t)=min⁡(t60000,1)h(t) = \min\left(\frac{t}{60000},1\right)

The depth-loss weight is

wd(t)=0.5(0.005)h(t)w_d(t) = 0.5(0.005)^{h(t)}

which decreases from 0.5 to 0.0025.

The Gaussian-shaping weight is zero for t<3000t < 3000 and otherwise follows

ws(t)=0.01(0.01)h(t),t≥3000w_s(t) = 0.01(0.01)^{h(t)}, \qquad t \geq 3000

The planarity weight is zero for t<6000t < 6000 and otherwise follows

wp(t)=0.2(0.05)h(t),t≥6000w_p(t) = 0.2(0.05)^{h(t)}, \qquad t \geq 6000

Thus, shaping begins with a weight of approximately 0.00794 and decays to 10−410^{-4}, while planarity begins at approximately 0.148 and decays to 0.01. The coarse stage uses only photometric and depth supervision.

Gaussian shaping penalizes the scale ratio smin⁡/smax⁡s_{\min}/s_{\max} for Gaussians whose ratio exceeds τs=0.2\tau_s=0.2. The soft axis-selector temperature is T=0.1T=0.1, and densification is performed from step 1k to step 30k. We will include these settings in the appendix for reproducibility.

We thank the reviewer again for the constructive comments. We believe the additional measurements and clarified mathematical and experimental scope more precisely establish the contributions, limitations, and computational tradeoffs of UrbanBP-GS.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.