acceptodds
Under review as a conference paper at ICLR 2027

PanoLOG: Geometry and Gradient-based Partitioning for Panoramic Outdoor Reconstruction

Abstract

Scaling 3D Gaussian Splatting (3DGS) to large outdoor scenes is costly in both data acquisition and computation. Adopting panoramic images with equirectangular projection (ERP) can reduce capture effort via their full 360° field of view, yet the resulting omnipresent visibility invalidates existing partitioning strategies that rely on local camera frustums, causing block-wise optimization to degenerate into global training. Thus, we propose PanoLOG (Panoramic Large-scale Outdoor Gaussian Splatting), a two-stage coarse-to-fine framework equipped with a Geometry and Gradient-based Partitioning Strategy (G²PS) tailored for large-scale panoramic 3DGS reconstruction. In the global coarse stage, PanoLOG leverages sky-sphere modeling and panoramic monocular depth supervision for reliable geometry, while in the refinement stage, G²PS builds adaptive bounding volumes via parallax-driven uncertainty and assigns cameras via gradient-based importance scoring. Furthermore, we construct Pano360, the first large-scale panoramic benchmark for outdoor scene reconstruction. Extensive experiments demonstrate that PanoLOG achieves state-of-the-art rendering quality while maintaining scalable, block-parallel training. Our models, training code, and dataset will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Outdoor Reconstruction, Geometry

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kGfqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents PanoLOG, a two-stage framework for large-scale outdoor reconstruction from panoramic images using 3D Gaussian Splatting. The method first performs global coarse reconstruction with explicit sky modeling and monocular depth supervision, followed by block-wise refinement using a Geometry and Gradient-based Partitioning Strategy (G²PS). G²PS constructs spatial boundaries based on camera geometry and a parallax-inspired uncertainty formulation, and assigns cameras to individual blocks using rendering-loss gradients. The authors also introduce Pano360, a panoramic outdoor reconstruction benchmark containing 5,637 images across four scenes. Experiments on Pano360 and two public panoramic datasets report competitive rendering quality and relatively compact models compared with existing Gaussian Splatting approaches.

Critical strengths

The paper addresses the challenge of spatial partitioning for large-scale panoramic Gaussian Splatting, where conventional frustum-based strategies are less effective.

The proposed Pano360 benchmark provides a useful evaluation setting, and the method achieves competitive reconstruction quality with relatively compact models.

Critical weaknesses
  1. The central gradient-based partitioning mechanism lacks sufficient mathematical and methodological justification. Equation (10) defines camera–block importance using the average magnitude of rendering-loss gradients with respect to Gaussian positions. However, gradient magnitude is not necessarily equivalent to observational contribution or geometric visibility. A Gaussian may be highly relevant to a viewpoint while exhibiting a small positional gradient when the rendering residual is already low. Conversely, a large gradient may arise from reconstruction errors or parameter sensitivity rather than meaningful observational importance. This distinction is particularly relevant because the importance scores are collected after global coarse optimization. Furthermore, Equation (11) does not guarantee sufficient camera coverage for each spatial block or overlapping camera assignments between neighboring blocks. Nevertheless, the paper relies on these properties to motivate effective block-wise supervision and cross-boundary consistency. The claimed relationship between gradient-based allocation and coherent reconstruction is therefore not adequately established by the mathematical formulation. The geometry-based component also raises concerns. Equations (7)–(9) motivate spatial boundary expansion through parallax-related depth uncertainty, but Appendix A fixes the triangulation range factor to 5 across all scenes without explaining how the underlying angular uncertainty or tolerance is estimated. Consequently, the practical method appears closer to a baseline-dependent heuristic than to the explicitly uncertainty-driven formulation emphasized in the paper. These concerns directly affect the technical validity and novelty of G²PS, which is presented as the paper's central contribution.

  2. The experiments do not adequately substantiate the claimed scalability or isolate the contribution of the proposed partitioning strategy. A central motivation of PanoLOG is to prevent panoramic block-wise reconstruction from degenerating into inefficient global optimization. However, Section 3.2.2 and Appendix D explicitly state that every block is initialized with the complete set of Gaussians from Stage I. Since the assigned cameras still provide full panoramic observations, it remains unclear how much of the global representation is rendered, differentiated, and optimized during refinement before opacity-based pruning and final spatial filtering. Restricting camera assignments and cropping the final Gaussian representation do not, by themselves, establish computationally localized optimization. Despite scalability being a central claim, the experiments do not report training time, peak GPU memory, total computational cost, or performance as the number of spatial blocks increases. Appendix A also describes separate global and block-wise training stages, while the experimental setup states that all methods use 30,000 iterations. The resulting total computational budgets are not sufficiently clarified. Consequently, the reported reductions in final model size cannot be interpreted as direct evidence of improved training scalability. The experimental attribution is also incomplete. On Pano360, competing large-scale methods use cubemap-converted perspective images, whereas PanoLOG employs native ERP rendering together with additional sky and depth supervision. Although component ablations are provided, the G²PS ablation is limited to a position-only partitioning comparison on one scene and does not separately quantify the contributions of geometry-based partitioning and gradient-based camera allocation. As a result, the current experiments demonstrate competitive reconstruction performance but do not convincingly establish that the proposed partitioning mechanism is responsible for the observed improvements or that it resolves the claimed scalability problem.

Questions for authors
  1. What evidence supports using positional rendering-loss gradients as a reliable measure of camera–block observational contribution? Equation (10) uses gradient magnitude as the importance criterion, although this quantity depends on both rendering residuals and parameter sensitivity. How do the authors distinguish genuinely informative observations from regions with large reconstruction errors? In particular, how does the criterion behave for well-reconstructed regions whose positional gradients become small after Stage I? Can the authors provide controlled evidence showing that gradient-based allocation outperforms alternative camera-assignment criteria under comparable camera budgets?

  2. Does Stage II actually achieve computationally localized optimization? The paper states that every block is initialized with the full Gaussian representation from Stage I. During block-wise refinement, are Gaussians outside the target block still rendered, updated, and considered for densification before opacity-based pruning? If so, how does the method avoid repeated optimization of substantial portions of the global scene? Can the authors report stage-wise training time, peak GPU memory, the number of actively optimized Gaussians, and computational scaling with the number of blocks? These measurements are essential for evaluating the paper's central scalability claim.

  3. How is the uncertainty-based geometric partitioning implemented and validated? Equations (7)–(9) derive the partitioning margin from angular parallax uncertainty, but Appendix A specifies a fixed triangulation range factor across all scenes. Are the angular uncertainty and tolerated relative depth error explicitly estimated, or is the range factor selected empirically? If the latter, what establishes that the proposed strategy is genuinely uncertainty-adaptive rather than a camera-baseline-scaled heuristic? Additionally, can the authors disentangle the effects of geometry-based boundary construction and gradient-based camera allocation through controlled ablations?

  4. Are the comparisons sufficient to attribute the reported improvements to G²PS under comparable computational budgets? The large-scale baselines operate on cubemap-converted images, while the proposed approach uses native ERP rendering and additional depth and sky modeling. How much of the reported improvement is attributable specifically to G²PS rather than these other differences? Were the total training costs, including both global optimization and independent block refinement, comparable across methods? A controlled comparison using consistent rendering and supervision configurations would materially affect the assessment of the proposed contribution.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer kGfq, Q1-part1

We thank the reviewer for their patience. This rebuttal involved a large number of additional experiments, which took some time to complete under limited GPU resources. We kindly invite the reviewer to read the additional results below, and we are happy to answer any further questions.

Reply to Q1 (Part 1/2):

In the paper, we state that the position gradient reflects the importance of a Gaussian to the current viewpoint: regions with large contributions exhibit large gradients, while distant or occluded regions have near-zero gradients. Here, "importance" means the strength of the optimization signal that the camera's photometric loss provides to the position parameters of the Gaussians in a block, not a direct estimate of physical visibility.

The gradient score of camera kk for the Gaussians in block bb is defined as

sk,b=1∣Gb∣∑g∈Gb∥∂Lk∂xg∥1,s_{k,b} = \frac{1}{|\mathcal{G}_b|}\sum_{g\in \mathcal{G}_b}\Big\|\frac{\partial \mathcal{L}_k}{\partial \mathbf{x}_g}\Big\|_1,

with the same notation as Eq. (10) of our paper. From the volume rendering C=∑iciαi∏j<i(1−αj)C=\sum_i c_i\alpha_i\prod_{j<i}(1-\alpha_j), we obtain

∂C∂αn=Tn(cn−Cn+),Tn=∏j<n(1−αj),Cn+=∑i>nciαi∏n<j<i(1−αj),\frac{\partial C}{\partial \alpha_n}=T_n(c_n-C_{n+}),\quad T_n=\prod_{j<n}(1-\alpha_j),\quad C_{n+}=\sum_{i>n}c_i\alpha_i\prod_{n<j<i}(1-\alpha_j),

where Cn+C_{n+} is the composited color of all Gaussians behind the nn-th Gaussian. Hence Tn>0T_n>0 (the Gaussian is not fully occluded) is a necessary condition for a non-zero gradient, since both terms contain TnT_n. It is not a sufficient condition: the gradient can be zero when cn=Cn+c_n=C_{n+} (the Gaussian has the same color as the composite behind it), or when the pixel residual is zero, which gives ∂Lk/∂C=0\partial\mathcal{L}_k/\partial C=0. Conversely, a non-zero gradient implies that the Gaussian projects onto the valid pixel region of the ERP image, is not fully occluded, and makes a non-zero contribution to the pixels.

We do not treat this score as an exact measure of visibility, but as the strength of the optimization signal that the current model still needs under that camera's observation. We therefore do not attempt to distinguish "genuinely informative observations" from "large reconstruction errors": both mean that the current model is still sensitive to that observation, and both should receive optimization signal. The only difference is that converged regions with small gradients do not need additional cameras.

On the concern that well-reconstructed regions have small residuals and hence small position gradients: Eq. (11) uses the score normalized across blocks for the same camera, sk,b/max⁡b′sk,b′s_{k,b}/\max_{b'}s_{k,b'}, not its absolute magnitude. When the overall gradient scale of a camera shrinks, its scores on all blocks shrink proportionally and the assignment is unchanged. The assignment Jaccard of 0.97 to 0.99 between C20K and C30K (Table 3) also indicates that the assignment is fairly stable within the tested range. More importantly, the geometric membership term of Eq. (11) guarantees that every camera is assigned at least to the block containing its center, so a well-reconstructed region still keeps all the cameras located inside it. The gradient term only determines which cameras from other blocks are additionally assigned, so small gradients affect only these "additional" cameras and do not deprive such regions of basic supervision.

On the coverage and overlap of Eq. (11): camera centers lie inside the AABB and the blocks form a partition of the AABB, so the geometric term leaves no camera unassigned and gives each block at least all the cameras located inside it. The gradient term then adds the cameras from other blocks whose normalized score exceeds τgrad\tau_{\mathrm{grad}}. With the default 2 blocks, the camera coverage of the two blocks is 40.29%/73.33% on NSC and 52.98%/59.92% on NSK. Moreover, every block in Stage II is initialized from, and optimizes, the complete set of Stage I Gaussians, and is filtered by its spatial extent only at the end, so Gaussians near a block boundary are optimized jointly under the camera sets of the adjacent blocks. Even when the camera overlap is small, the shared initialization and two-sided observation help maintain cross-boundary consistency.

AuthorsOpenReview ↗

Responses to Reviewer kGfq, Q1-part2

Reply to Q1 (Part 2/2):

To validate the gradient-based assignment under the same camera budget, we match the number of cameras per block of only-position to that of G²PS: cameras outside a block are added from the block boundary outward, nearest first, until the block has the same number of cameras as under G²PS. The Stage I initialization, partition, ERP rendering, and supervision are identical.

Table 1: only-position with matched camera budget vs. G²PS.

SceneSettingCameras per BlockPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSConly-position (camera-matched)40.29% / 73.33%27.690.8450.258458
NSCG²PS40.29% / 73.33%28.140.8590.246426
NSKonly-position (camera-matched)52.98% / 59.92%24.150.8120.198623
NSKG²PS52.98% / 59.92%24.640.8270.187597

Here, the model size is an outcome of training rather than a controlled variable. The capacity-controlled comparison with only-position (without camera matching) and the full ablation of only-position / only-grad / G²PS are given in Table 7 in our reply to Q4.

We further compare with the rendering contribution based assignment. CityGaussian estimates the contribution of a camera to a block by removing that block and measuring the SSIM change of the rendered image. For ERP panoramas, a single camera has a 360° observation range, so different blocks all contribute to the whole image, and this measure may assign many cameras to multiple blocks. On NSC and NSK, we measured the fraction of all training cameras that each method assigns to each block.

Table 2: Camera assignment ratios of the SSIM-based and gradient-based assignments.

DatasetBlockSSIM AssignmentGradient Assignment
NSCBlock 1100%40.29%
NSCBlock 2100%73.33%
NSKBlock 1100%52.98%
NSKBlock 2100%59.92%

The SSIM-based assignment gives all training cameras to both blocks, whereas the gradient-based assignment produces more selective camera sets. Since the SSIM-based assignment loses its discriminative power, we did not proceed to Stage II training under it.

On the concern that gradients may become small after Stage I converges: we compute the gradient scores and the camera assignment from the C20K and C30K checkpoints (20K and 30K iterations) of the same Stage I run. To avoid the influence of the initialization quality of the coarse model, Stage II is always initialized from C30K; only the checkpoint used to compute the gradient score changes. The reported Jaccard is the Jaccard similarity of the camera sets under the two checkpoints, computed per block and averaged over blocks.

Table 3: Sensitivity to the coarse checkpoint.

DatasetCoarse CheckpointJaccard ↑PSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSCC20K0.9928.090.8540.248394
NSCC30K128.140.8590.246426
NSKC20K0.9724.580.8260.189572
NSKC30K124.640.8270.187597

Between C20K and C30K, the camera-block assignment Jaccard reaches 0.99 on NSC and 0.97 on NSK, and the final PSNR differs by only 0.05 dB and 0.06 dB, respectively. Within the tested late training range, the gradient-guided camera assignment is therefore highly stable and does not change significantly as the coarse model is further optimized.

AuthorsOpenReview ↗

Responses to Reviewer kGfq, Q2

Reply to Q2:

We first clarify how Stage II is actually trained.

Every block is initialized from the complete Stage I coarse Gaussian model, but is refined only with the training cameras assigned to it. Since each ERP camera observes the full 360° environment, we keep the Gaussians outside the target block in the panoramic rendering and parameter updates instead of removing them before training. These Gaussians are optimized as well; whether they are updated and densified depends on their visibility and gradients under the cameras assigned to the current block.

This design keeps the panoramic rendering context complete. Removing the Gaussians outside the target block in advance would leave the other regions observed by the current camera without a representation, which could impair the supervision from the full panorama. After training, we keep only the Gaussians inside the spatial extent of each block and merge them into the final model.

We note that this process is not equivalent to strictly local Gaussian optimization. The Gaussians outside the target block still participate in training, so our method does not fully eliminate the repeated rendering and parameter optimization cost across blocks. The partitioning strategy mainly organizes the refinement of different regions through the assignment of camera sets, and the final spatial filtering avoids keeping duplicate copies of the full scene representation.

On computational efficiency, we report on NSK the training time, peak GPU memory, and number of active Gaussians with different numbers of blocks, together with the stage-wise computational cost. Coarse GS is the total number of Gaussians in the Stage I coarse model. Active GS is the number of Gaussians that, when a block loads the coarse model, are visible to the cameras assigned to that block and satisfy the densification criterion.

Table 4: Computational efficiency with different numbers of blocks.

#BlocksBlockCoarse GS (M)Train Time (min)Peak Mem (GB)Active GS (M)
201.9335.97.61.23
211.9336.77.81.02
401.9338.15.20.88
411.9342.85.40.90
421.9340.05.00.78
431.9336.35.90.56

Table 5: Stage-wise computational cost (NSK, 2 blocks).

StageTime (min)Peak Mem (GB)
Stage I3911.8
Partition83.2
Stage II72.67.8
Merging0.943.6
Total120.54

These results show the actual training mechanism of PanoLOG, its computational cost, and how the number of blocks affects resource consumption.

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

This paper presents PanoLOG, a coarse-to-fine 3D Gaussian Splatting framework for large-scale outdoor panoramic scene reconstruction. Its core component, G²PS, constructs spatial partitions using parallax-driven geometric constraints and adaptively assigns training cameras to different spatial blocks based on the gradients of the rendering loss with respect to Gaussian positions. The framework further incorporates explicit sky-sphere modeling and panoramic monocular depth supervision. Experimental results demonstrate that PanoLOG achieves competitive rendering quality on the newly introduced Pano360 benchmark and several existing panoramic datasets.

Critical strengths
  1. Practically relevant research problem. This paper addresses the challenges that the 360° field of view of panoramic cameras poses to conventional spatial partitioning and training-view assignment strategies in large-scale 3DGS. The problem is meaningful and has practical relevance.

  2. Simple and straightforward method design. The proposed gradient-based camera assignment strategy does not require additional learning networks. It is intuitive, reasonably interpretable, and can be readily integrated into existing optimization-based 3DGS reconstruction pipelines.

  3. Competitive experimental results. The authors report favorable PSNR, SSIM, and LPIPS results across multiple datasets. Additionally, the newly introduced Pano360 dataset provides a potentially valuable benchmark for large-scale real-world outdoor panoramic reconstruction.

Critical weaknesses
  1. Limited novelty and insufficient differentiation from prior work.

The proposed framework shares substantial similarities with CityGaussian. In particular, CityGaussian already introduces an adaptive camera assignment strategy based on rendering contribution, rather than relying solely on camera frustum intersection or spatial proximity. Therefore, the proposed gradient-based scoring mechanism appears to be an alternative importance metric within an existing framework rather than a fundamentally new partitioning paradigm. The current experiments only compare G²PS against a position-based assignment strategy, which is insufficient to establish the advantages of gradient-based scoring over existing contribution-based approaches. I encourage the authors to provide a direct comparison with CityGaussian's original SSIM-based camera assignment strategy and further analyze the differences and applicability of these two importance metrics.

  1. Insufficient evaluation of training efficiency and scalability.

The paper repeatedly emphasizes scalable reconstruction and block-parallel training. However, the experiments primarily report rendering quality and final model size, without adequately evaluating training time, peak GPU memory consumption, or how computational costs scale with the number of partitions. Furthermore, the main experiments are conducted on a single NVIDIA RTX 4090 GPU (24 GB), which does not directly demonstrate the benefits of multi-GPU parallel training. I suggest that the authors report training time, memory consumption, camera assignment statistics, and Gaussian counts under different numbers of partitions to better substantiate the scalability claims.

  1. Insufficient validation of the adaptivity of parallax-based spatial boundary estimation.

The paper derives a triangulation range factor through triangulation error propagation:

ρtri=ησγ\rho_{\mathrm{tri}} = \frac{\eta}{\sigma_\gamma}

which is subsequently used to determine the spatial boundary expansion. However, in the experiments, this factor is simply fixed to ρtri=5\rho_{\mathrm{tri}}=5, without demonstrating how it is estimated from the actual parallax distribution or triangulation uncertainty of different scenes. Consequently, the claimed adaptivity of the proposed geometric partitioning strategy is not sufficiently substantiated.

  1. Confounding effects of model capacity in the ablation study.

According to Table 4, the full model achieves a PSNR improvement of approximately 0.51 dB over the variant without G²PS. However, this improvement is accompanied by an increase in model size from 580 MB to 667 MB, corresponding to approximately 15% additional model capacity. Since a larger model may itself improve reconstruction quality, the current results do not sufficiently disentangle the benefits of more effective camera assignment from those of increased Gaussian representation capacity. I suggest that the authors conduct controlled experiments with matched Gaussian counts or comparable model sizes to more convincingly demonstrate the independent contribution of G²PS.

Questions for authors

Seen above.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer cyxF, Q1

We thank the reviewer for their patience. This rebuttal involved a large number of additional experiments, which took some time to complete under limited GPU resources. We kindly invite the reviewer to read the additional results below, and we are happy to answer any further questions.

Reply to Q1 (Weakness 1):

We agree that our work does not propose an entirely new block-wise reconstruction paradigm independent of prior work. Targeting the 360° observation of ERP panoramas, we design a spatial partitioning and camera assignment strategy that combines geometric boundaries with gradient guidance.

CityGaussian uses a block-removal rendering contribution measure: the SSIM difference between the images rendered before and after removing the target block. G²PS instead uses the gradient magnitude of a single camera's photometric loss with respect to the positions of the Gaussians in a block, which measures the strength of the optimization signal that the camera provides to that block. The two criteria thus differ in what they measure: the effect of removing a spatial region on the current rendering, versus the optimization signal that the current observation provides to the geometric parameters in that region.

For ERP panoramas, a single camera has a full 360° observation range, so different blocks all contribute to the same panorama. Assigning cameras by the whole-image SSIM difference may therefore assign many cameras to multiple blocks and weaken the spatial selectivity of block-wise training.

We verified this on NSC and NSK by measuring the fraction of all training cameras assigned to each block. The SSIM-based assignment uses the official CityGaussian implementation with its default rendering contribution (block-removal SSIM) rule, applied to the same two blocks, Stage I initialization, and ERP renderer.

Table 1: Camera assignment ratios of the SSIM-based and gradient-based assignments.

DatasetBlockSSIM AssignmentGradient Assignment
NSCBlock 1100%40.29%
NSCBlock 2100%73.33%
NSKBlock 1100%52.98%
NSKBlock 2100%59.92%

The SSIM-based assignment gives all training cameras to both blocks. The camera sets of the blocks then overlap completely, so the assignment loses its discriminative power, and GPU memory becomes insufficient on large scenes. We therefore did not proceed to Stage II training under this assignment.

In addition, under the same ERP rendering and supervision and with model sizes kept as close as possible, we compare only-position, only-grad, and the full G²PS. Here, only-position keeps only the geometric membership term of Eq. (11) (a camera is assigned to the block containing its center), only-grad keeps only the gradient term, and the full G²PS uses both.

Table 2: Ablation of partitioning and camera assignment strategies.

SceneSettingPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSConly-position27.800.8510.252406
NSConly-grad27.840.8510.251412
NSCG²PS28.140.8590.246426
NSKonly-position24.080.8160.193614
NSKonly-grad24.380.8210.193609
NSKG²PS24.640.8270.187597

The full G²PS achieves higher PSNR and SSIM and lower LPIPS on both scenes, which further indicates that combining geometric boundary construction with gradient-guided assignment improves reconstruction quality under the current setting.

AuthorsOpenReview ↗

Responses to Reviewer cyxF, Q2

Reply to Q2 (Weakness 2):

We agree that our submission focused on reconstruction quality and final model size, and did not sufficiently report the computational cost of the full pipeline or the resource consumption with different numbers of blocks.

We therefore tested 2, 4, and 8 blocks on NSK, NSN, and BAX on an RTX 4090. Overlap is the average number of blocks to which each training image is assigned. Total Time is the sum of the Stage I, Partition, Stage II, and Merging times. Stage II Time is the cumulative time of training the blocks sequentially on a single GPU. The refinement of each block is independent and can run in parallel on multiple GPUs; we did not conduct multi-GPU experiments in this work.

Table 3: Reconstruction quality and computational cost with different numbers of blocks.

NSK

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
224.650.8270.1861.137.81.212.01
424.320.8310.1771.225.92.623.42
823.290.8230.1791.324.85.466.26

NSN

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
224.610.7520.3281.029.81.933.04
424.750.7700.3051.057.63.965.07
823.930.7640.3201.125.97.628.73

BAX

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
221.350.6660.4171.028.21.812.85
421.460.6630.4161.066.33.844.88
820.490.6560.4201.075.27.348.38

As the number of blocks increases, the cumulative training time grows markedly, while reconstruction quality does not keep improving. Across the three scenes, 2 to 4 blocks generally give a better trade-off between quality and computational cost, and PSNR drops to some extent with 8 blocks. The Stage II peak memory decreases as the number of blocks increases.

These results indicate that the partition granularity needs to match the scene scale, camera distribution, and training budget. We believe that an overly fine partition may aggravate the imbalance of local supervision and hinder cross-block optimization.

On NSK, we also report the number of Gaussians, training time, and peak GPU memory of each block. Coarse GS is the total number of Gaussians in the Stage I coarse model. Active GS is the number of Gaussians that, when a block loads the coarse model, are visible to the cameras assigned to that block and satisfy the densification criterion.

Table 4: Per-block Gaussian counts and computational cost on NSK with different numbers of blocks.

#BlocksBlockCoarse GS (M)Train Time (min)Peak Mem (GB)Active GS (M)
201.9335.97.61.23
211.9336.77.81.02
401.9338.15.20.88
411.9342.85.40.90
421.9340.05.00.78
431.9336.35.90.56
AuthorsOpenReview ↗

Responses to Reviewer cyxF, Q3 and Q4

Reply to Q3 (Weakness 3):

We clarify the relation between the theoretical derivation of the geometric boundary and the actual parameter setting.

We use the parallax error propagation of triangulation as the geometric basis of the spatial boundary expansion. The expansion distance is determined jointly by the representative camera baseline and the triangulation range factor:

m=b^⋅ρtri,m=\hat b\cdot\rho_{\mathrm{tri}},

where b^\hat b is the median nearest-neighbor camera distance and ρtri\rho_{\mathrm{tri}} is the triangulation range factor.

In large outdoor panoramic scenes, the 360° observation covers both near and far content, so the initial sparse point cloud may contain triangulated points far from the camera trajectory. Building the core partitioning range from the extrema of the whole point cloud would let these distant points significantly enlarge the AABB and unbalance the spatial partition. We therefore build the core boundary from the camera trajectory and control the expansion distance with the representative camera baseline and the triangulation range factor, so that the extent of the core region varies with the camera baseline scale of each scene.

In our current implementation, ρtri\rho_{\mathrm{tri}} is fixed to 5 as an empirically chosen hyperparameter; η\eta and σγ\sigma_\gamma are not estimated explicitly and separately. The scene adaptivity of our method comes from b^\hat b: the median nearest-neighbor baseline differs across scenes, so the actual expansion distance varies with the scene scale.

We add a controlled experiment with different values of ρtri\rho_{\mathrm{tri}}. Within the tested range (2 to 10), changing ρtri\rho_{\mathrm{tri}} has little effect on the overall reconstruction quality. With the default number of blocks, reconstruction quality is insensitive to the exact expansion distance, and the main role of the geometric term is to stabilize the extent of the AABB.

Table 5: Sensitivity analysis of the triangulation factor.

Datasetρtri\rho_{\mathrm{tri}}PSNR ↑SSIM ↑LPIPS ↓
NSC228.1420.8580.246
NSC528.1830.8590.243
NSC1028.0980.8530.245
NSK224.6170.8240.190
NSK524.6430.8270.187
NSK1024.6230.8240.189

Reply to Q4 (Weakness 4):

We agree that, in the original ablation, the full G²PS and the variant without G²PS differ in the final number of Gaussians and model size, so the effect of increased model capacity on reconstruction quality could not be fully ruled out.

We therefore retrained a capacity-controlled comparison. The Stage I initialization, training data, optimization settings, and number of iterations are identical; we adjust only the densification settings so that the final model sizes are close (NSC: 406 vs 426 MB; NSK: 614 vs 597 MB), which reduces the effect of differences in Gaussian representation capacity.

Table 6: Capacity-controlled ablation of G²PS.

SceneSettingPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSConly-position (w/o G²PS)27.800.8510.252406
NSCG²PS28.140.8590.246426
NSKonly-position (w/o G²PS)24.080.8160.193614
NSKG²PS24.640.8270.187597

On NSC, the full G²PS improves PSNR by 0.34 dB over the variant without G²PS, with the model size increasing by only 20 MB (4.9%). On NSK, its model is 17 MB (2.8%) smaller, yet PSNR still improves by 0.56 dB, SSIM increases from 0.816 to 0.827, and LPIPS decreases from 0.193 to 0.187. With comparable model sizes, the full G²PS thus still achieves higher reconstruction quality. On NSK in particular, the model is smaller and the quality is higher, which indicates that the improvement does not come from increased model capacity.

AuthorsOpenReview ↗

Revised Paper Uploaded: Summary of Changes

Dear Reviewer,

We thank all reviewers for their comments and suggestions. In addition to our individual responses, we have incorporated the concerns shared by the reviewers into the main text of the revised paper, with all added content marked in blue. We appreciate this discussion and are happy to answer any further questions promptly.

The main text is now 10 pages. The changes are as follows.

  • Sec. 3.2.2: two clarifying sentences, stating that ρtri\rho_{\mathrm{tri}} is fixed to 5 as an empirically chosen hyperparameter, and that the gradient score measures the strength of the optimization signal rather than physical visibility.
  • Sec. 4.1: clarification of the training budget (30,000 iterations per stage and per block).
  • Sec. 4.3: a capacity-controlled ablation of G²PS on NSC and NSK (Table 6), and a comparison with the rendering-contribution assignment of CityGaussian.
  • Sec. 4.5 (new): the effect of block count on quality and training cost (Table 7).

Best regards, The Authors

Reviewer_TkF6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PanoLOG reconstructs large outdoor scenes from panoramas using two-stage Gaussian Splatting. Global training combines an ERP renderer, sky modeling, and monocular depth supervision. G²PS then partitions space and assigns cameras using geometry and position gradients before independent block refinement and merging. The paper introduces Pano360 and reports competitive rendering quality on this dataset, Ricoh360, and 360Roam. Computational scalability and the specific contribution of G²PS remain insufficiently validated.

Critical strengths
  1. Encouraging reconstruction quality. PanoLOG improves PSNR over the strongest reported baselines by 0.64/1.16 dB on BAX/NSN, with substantially smaller models than H3DGS.
  2. Potential benchmark value. Pano360 provides 5,637 aerial and handheld panoramas. Evaluation on two public datasets supports applicability beyond the authors’ benchmark.
Critical weaknesses
  1. Insufficient isolation of the central contribution. CityGaussian already uses rendering-contribution-based camera selection, so panoramic visibility alone does not establish its failure. The single-scene, position-only ablation neither compares this closer alternative nor separates geometry from gradient allocation. Loss gradients also reflect reconstruction errors rather than visibility alone.
  2. Unverified scalability and ambiguous budgets. Each block loads the complete coarse model, but training time, peak memory, and scaling results are absent. The stated 30,000-iteration budget is ambiguous across stages and blocks. Missing panoramic baselines on Pano360 and a matched-budget global ERP control further weaken attribution.
Questions for authors
  1. How are training budgets matched across stages, blocks, and cubemap versus panoramic inputs?
  2. How do runtime, peak memory, and camera overlap change with block count?
  3. Under identical rendering and supervision, how does G²PS compare with rendering-contribution allocation and continued global training?
  4. Can geometry and gradient allocation be ablated separately, including sensitivity to the coarse checkpoint and triangulation factor?
  5. How do panoramic baselines perform on Pano360 under comparable budgets?
5 comments
AuthorsOpenReview ↗

Responses to Reviewer TkF6, Q1 and Q2

We thank the reviewer for their patience. This rebuttal involved a large number of additional experiments, which took some time to complete under limited GPU resources. We kindly invite the reviewer to read the additional results below, and we are happy to answer any further questions.

Reply to Q1:

We agree that the training budget was not stated clearly in our submission and clarify it here.

The 30,000 iterations stated in our submission are the budget of each optimization stage or spatial block (partition), not the total budget of the pipeline. All methods follow the same protocol: 30,000 iterations of global/coarse training, followed by 30,000 refinement iterations per block. For the same number of blocks, the total budget is therefore identical across methods: 30k + 2×30k = 90k with 2 blocks. PanoLOG uses 2 blocks by default (90k), and 150k and 270k with 4 and 8 blocks. H3DGS uses 8 blocks on BAX/NSN, i.e., 270k in total, which is 3× the default budget of PanoLOG.


Reply to Q2:

Following the reviewer's suggestion, we tested 2, 4, and 8 blocks on NSK, NSN, and BAX on an RTX 4090. Overlap is the average number of blocks to which each training image is assigned. Total Time is the sum of the Stage I, Partition, Stage II, and Merging times. Stage II Time is the cumulative time of training the blocks sequentially on a single GPU.

Table 1: Reconstruction quality and computational cost with different numbers of blocks.

NSK

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
224.650.8270.1861.137.81.212.01
424.320.8310.1771.225.92.623.42
823.290.8230.1791.324.85.466.26

NSN

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
224.610.7520.3281.029.81.933.04
424.750.7700.3051.057.63.965.07
823.930.7640.3201.125.97.628.73

BAX

#BlocksPSNR ↑SSIM ↑LPIPS ↓OverlapStage II Peak Mem (GB)Stage II Time (h)Total Time (h)
221.350.6660.4171.028.21.812.85
421.460.6630.4161.066.33.844.88
820.490.6560.4201.075.27.348.38

As the number of blocks increases, the cumulative Stage II time and the total time increase markedly because each block requires its own refinement. The camera assignment overlap also increases, while the Stage II peak memory decreases.

The relation between the number of blocks and reconstruction quality is not monotonic. On NSK, 2 blocks give the highest PSNR and 4 blocks give better SSIM and LPIPS; on NSN and BAX, 4 blocks are better on some metrics. With 8 blocks, PSNR drops on all three scenes.

These results indicate that the partition granularity needs to match the scene scale, camera distribution, and optimization budget. An overly fine partition may aggravate the imbalance of local supervision and hinder cross-block consistency. We do not claim that more blocks necessarily improve reconstruction quality or reduce the total training cost; an appropriate granularity gives a better trade-off between quality and cost.

AuthorsOpenReview ↗

Responses to Reviewer TkF6, Q3

Reply to Q3:

We clarify the difference between G²PS and existing rendering contribution strategies.

Our core contribution is not a fundamentally new block-wise reconstruction paradigm, but a geometric partitioning and camera assignment strategy suited to the 360° field of view of ERP panoramas.

CityGaussian scores the contribution of a camera to a block by block-removal SSIM, i.e., the difference between the images rendered from the complete Gaussian model and from the model without the target block, combined with the camera position. With perspective images, the limited field of view keeps this contribution spatially local. An ERP panorama observes multiple blocks at once, so removing any of them noticeably changes the whole image. The whole-image SSIM difference may therefore assign many panoramas to multiple blocks and weaken the discriminative power of the assignment.

We verified this on NSC and NSK by measuring the fraction of all training cameras assigned to each block. The SSIM-based assignment uses the official CityGaussian implementation with its default rendering contribution (block-removal SSIM) rule, applied to the same two blocks, Stage I initialization, and ERP renderer.

Table 2: Camera assignment ratios of the SSIM-based and gradient-based assignments.

DatasetBlockSSIM AssignmentGradient Assignment
NSCBlock 1100%40.29%
NSCBlock 2100%73.33%
NSKBlock 1100%52.98%
NSKBlock 2100%59.92%

The SSIM-based assignment gives 100% of the training cameras to both blocks, whereas the gradient-based assignment gives 40.29% and 73.33% on NSC and 52.98% and 59.92% on NSK. With ERP inputs, the SSIM-based rule thus degenerates to "every block uses all training cameras" and loses its discriminative power, so we did not proceed to Stage II training under this assignment.

To isolate the contribution of G²PS under identical rendering, supervision, and budget, we compare only-position, only-grad, and the full G²PS (Table 3 in our reply to Q4). Compared with only-position, the full G²PS improves PSNR by 0.34 dB on NSC and 0.56 dB on NSK, with a smaller model on NSK (597 MB vs 614 MB).

On the concern that loss gradients also reflect reconstruction errors: we interpret the gradient score not as visibility, but as the strength of the optimization signal that the current model still needs under that camera's observation. The geometric membership term of Eq. (11) guarantees that every camera is assigned at least to the block containing its center; the gradient term only determines which cameras are additionally assigned.

AuthorsOpenReview ↗

Responses to Reviewer TkF6, Q4

Reply to Q4:

Following the reviewer's suggestion, we add an ablation of the camera assignment strategy and sensitivity analyses on the coarse checkpoint and the triangulation factor.

(1) Separate ablation of geometry and gradient

We compare three configurations. Here, only-position keeps only the geometric membership term of Eq. (11) (a camera is assigned to the block containing its center), only-grad keeps only the gradient term (assignment by the normalized gradient score), and the full G²PS uses both. The three configurations share the AABB boundary construction, Stage I initialization, number of blocks (2), training images, ERP rendering, and depth/sky supervision, with model sizes kept as close as possible. "Geometry" refers to the geometric membership term of Eq. (11); the sensitivity to the AABB boundary parameter (ρtri\rho_{\mathrm{tri}}) is analyzed in part (3).

Table 3: Ablation of partitioning and assignment strategies.

SceneSettingPSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSConly-position27.800.8510.252406
NSConly-grad27.840.8510.251412
NSCG²PS28.140.8590.246426
NSKonly-position24.080.8160.193614
NSKonly-grad24.380.8210.193609
NSKG²PS24.640.8270.187597

The full G²PS achieves the highest PSNR and SSIM and the lowest LPIPS on both NSC and NSK.

(2) Sensitivity to the coarse checkpoint

The gradient-based assignment relies on the Stage I coarse model to compute position gradients. We therefore compute the camera-block assignment from the C20K and C30K checkpoints (20K and 30K iterations) of the same Stage I run. To exclude differences in initialization quality, Stage II is always initialized from C30K; only the checkpoint used for gradient scoring, and hence the camera assignment, changes. We measure consistency by the Jaccard similarity of the camera sets under the two checkpoints, computed per block and averaged over blocks.

Table 4: Sensitivity to the coarse checkpoint.

DatasetCoarse CheckpointJaccard ↑PSNR ↑SSIM ↑LPIPS ↓Size (MB) ↓
NSCC20K0.9928.090.8540.248394
NSCC30K128.140.8590.246426
NSKC20K0.9724.580.8260.189572
NSKC30K124.640.8270.187597

C30K is the reference checkpoint, so its Jaccard with itself is 1. The assignments from C20K and C30K are highly consistent, with a Jaccard similarity of 0.99 on NSC and 0.97 on NSK, and the final PSNR differs by only 0.05 dB and 0.06 dB, respectively.

(3) Sensitivity to the triangulation factor

We compute the expansion margin from the representative baseline b^\hat b of the camera trajectory and the triangulation range factor ρtri\rho_{\mathrm{tri}}:

m=b^⋅ρtri.m=\hat b\cdot\rho_{\mathrm{tri}}.

Our current implementation uses a fixed triangulation range factor, while the actual expansion distance varies with the camera baseline scale of each scene. We test ρtri=2,5,10\rho_{\mathrm{tri}}=2,5,10.

Table 5: Sensitivity analysis of the triangulation factor.

Datasetρtri\rho_{\mathrm{tri}}PSNR ↑SSIM ↑LPIPS ↓
NSC228.1420.8580.246
NSC528.1830.8590.243
NSC1028.0980.8530.245
NSK224.6170.8240.190
NSK524.6430.8270.187
NSK1024.6230.8240.189

The maximum PSNR difference is 0.085 dB on NSC and 0.026 dB on NSK, indicating that this geometric expansion parameter is fairly stable within a reasonable range.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.