{"id":"4b18b030-f390-4fae-9fd7-dd7ece716441","arxiv_id":"2607.08769","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"A geometry and gradient-based partitioning strategy enables scalable block-wise 3D Gaussian Splatting for large-scale panoramic outdoor scenes.","lead":"This paper introduces a method to split large outdoor 360-degree panoramic scenes into smaller blocks for efficient 3D reconstruction, solving the problem that panoramic cameras see everything at once. It matters because it makes building high-quality 3D maps of large outdoor areas cheaper and more computationally feasible.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"G2PS never reports camera-per-block reduction ratios; without this, the core claim that partitioning prevents degeneration to global training is unverified.","rationale":"The reader correctly identified gradient reliability as a concern, but the more fundamental issue is that the paper never demonstrates the partitioning achieves its stated goal. The headline claim has two parts: (1) quality and (2) scalable block-parallel training. The quality results are reasonably supported by Tables 1-3, though the baseline comparison is somewhat confounded by cubemap conversion of pinhole methods. The scalability claim, however, rests entirely on G2PS actually reducing per-block camera counts and memory — and this is never measured or reported. The ablation (Table 4) shows G2PS contributes only 0.51 dB, comparable to the sky sphere's 0.50 dB, which raises the question of whether the partitioning is doing meaningful work or whether the gains come from auxiliary components. The paper promises code and data release, which would allow verification, but the manuscript itself does not provide the key evidence. A CONDITIONAL verdict is appropriate: accept if the authors can show that camera-per-block ratios are meaningfully below 100% (ideally <50%) and that per-block memory is substantially lower than global training. The method is well-motivated and the ERP-native design is sound, but the core partitioning claim needs direct verification rather than being inferred from end-to-end quality metrics.","tokens_in":12901,"tokens_out":2225,"duration_ms":78981,"concrete_test":"For each scene in Pano360, report the average number of cameras assigned per block (via Eq. 11) as a fraction of total cameras, and the average GPU memory per block during Stage II training. If the average fraction exceeds ~60% or memory per block is close to global training memory, the partitioning is not achieving meaningful locality and the scalability claim weakens significantly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that G2PS resolves the omnipresent visibility problem, enabling scalable block-wise training rather than degenerating into global optimization. But the paper never reports how many cameras are actually assigned per block relative to the total camera count. Equation (11) assigns camera k to block b if either (a) the camera is geometrically inside the block, OR (b) the normalized gradient score exceeds τ_grad=0.8. In panoramic scenes where every camera observes the full 360° environment, a camera may have significant gradient contributions to many blocks simultaneously. If the gradient score exceeds 0.8×max for, say, 5 out of 8 blocks, then each block still receives the majority of cameras, and the 'block-wise' training is barely distinguishable from global training. The ablation in Table 4 shows only a 0.51 dB drop when G2PS is removed entirely (replaced by position-only partitioning), which is surprisingly small if G2PS is truly preventing degeneration. This modest gain is consistent with the hypothesis that the partitioning provides marginal benefit over global training, and the main quality improvements come from the sky sphere (+0.50 dB), depth supervision, and ERP-native rendering rather than from the partitioning strategy itself. Without camera-per-block statistics, the claim that G2PS enables 'scalable, block-parallel training' is not substantiated — the method could be doing near-global training with minor pruning.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript presents PanoLOG, a two-stage coarse-to-fine framework for large-scale outdoor 3D Gaussian Splatting (3DGS) using panoramic (equirectangular projection, ERP) images. The core contribution is G2PS, a Geometry and Gradient-based Partitioning Strategy that addresses the 'omnipresent visibility' problem: because 360-degree cameras observe all directions, conventional frustum-based scene partitioning degenerates into global training. G2PS constructs an adaptive bounding volume via parallax-driven uncertainty and assigns cameras to spatial blocks using gradient-based importance scoring. The paper also introduces Pano360, a new large-scale panoramic outdoor dataset (5,637 images across four scenes). Experiments on Pano360 and public benchmarks (Ricoh360, 360Roam) show state-of-the-art rendering quality with smaller model sizes compared to baselines like H3DGS and CityGaussian.","tokens_in":13209,"tokens_out":1351,"duration_ms":151243,"significance":"The paper addresses a genuine technical gap: existing large-scale 3DGS partitioning methods rely on pinhole camera frustums, which are ill-suited for panoramic images. The proposed gradient-based camera-block allocation is a reasonable and novel mechanism to resolve this mismatch. The release of Pano360, a large-scale panoramic benchmark, and the public availability of code/models/dataset are significant assets that enhance reproducibility and enable future benchmarking. The ERP projection derivations (Appendix B) are mathematically sound. The experimental gains over strong baselines are notable, particularly the model size reductions (2.2-7.5x smaller than H3DGS).","major_comments":[{"comment":"The central claim that G2PS enables 'scalable, block-parallel training' and prevents degeneration to global training is not fully substantiated by the reported experiments. The paper does not report camera-per-block statistics (e.g., average number of cameras assigned per block relative to total camera count). In panoramic scenes, a camera may have significant gradient contributions to many blocks simultaneously. If the gradient score (Eq. 11) exceeds the threshold for multiple blocks, each block could still receive the majority of cameras, making the 'block-wise' training barely distinguishable from global training. The authors should report these statistics to verify that the partitioning actually reduces per-block camera counts. (§3.2.2, Eq. 11)","section":null},{"comment":"The ablation study (Table 4) shows only a 0.51 dB PSNR drop when G2PS is removed entirely (replaced by position-only partitioning), which is surprisingly small given that G2PS is the central contribution. This modest gain is consistent with the hypothesis that the main quality improvements come from the sky sphere, depth supervision, and ERP-native rendering rather than the partitioning strategy itself. The authors should provide a more detailed analysis justifying the necessity of G2PS for scalability (e.g., memory usage, training time, or performance on larger scenes where global training is infeasible). (Table 4, §4.3)","section":null},{"comment":"The gradient-based camera-block allocation (Eq. 11) assumes that the rendering loss gradient with respect to Gaussian positions is a reliable proxy for a camera's actual contribution to a spatial block. However, gradients can be noisy or influenced by factors unrelated to geometric contribution (e.g., texture density, lighting). The paper does not provide analysis or evidence that the gradient scores correlate with true geometric contribution. A visualization or quantitative analysis of the gradient scores would strengthen the validity of this assumption. (§3.2.2, Eq. 10-11)","section":null}],"minor_comments":[{"comment":"The threshold τ_grad = 0.8 is a hand-tuned hyperparameter. While Table 5 provides a sensitivity analysis, it would be beneficial to discuss how this value might need adjustment for scenes with different scales or camera densities. (Table 5, §4.4)","section":null},{"comment":"In Table 1, the model sizes for baselines vary significantly (e.g., H3DGS 1002.1 MB vs. CityGaussian 523.7 MB on NSC). It would be helpful to clarify whether all baselines were trained under comparable computational budgets or if differences in model size reflect methodological choices. (Table 1, §4.1)","section":null},{"comment":"The paper mentions that H3DGS uses eight partitions on BAX and NSN due to GPU memory limitations (§4.1), but it is unclear how many partitions PanoLOG uses on these scenes. This information should be explicitly stated for fair comparison. (§4.1)","section":null},{"comment":"Figure 3 is mentioned to show qualitative comparisons, but the resolution and clarity of the figure in the PDF make it difficult to discern the claimed improvements in distant regions and glass facades. Higher-quality figures would aid evaluation. (Figure 3)","section":null},{"comment":"The depth weight schedule (Eq. 17) uses an exponential decay from w_0 = 0.5 to w_T = 0.01. The rationale for these specific values is not discussed. A brief justification or reference to prior work would be helpful. (Appendix C, Eq. 17)","section":null},{"comment":"The paper states that 'sky Gaussian position gradients are zeroed' during Stage I (§3.2.1), but it is unclear whether this affects the optimization of appearance parameters (color, rotation, opacity) for sky Gaussians. The text later clarifies that appearance parameters are optimized normally, but this could be stated more clearly upfront. (§3.2.1)","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern regarding the lack of camera-per-block statistics is well-founded and is the primary reason for the major revision recommendation. The paper's central claim hinges on G2PS preventing degeneration to global training, but without evidence that the partitioning actually reduces per-block camera counts, this claim remains unverified. The modest ablation gain (0.51 dB) further weakens the case. If the authors can provide these statistics and demonstrate that G2PS genuinely enables scalable training (e.g., on scenes too large for global training), the paper would be significantly strengthened."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper you should know about is PanoLOG, which tackles a real problem: existing block-wise 3DGS partitioning strategies all assume pinhole frustum visibility, so they break down on 360° panoramic inputs where every camera sees everything. The core idea is G2PS — use parallax-driven geometry to build adaptive bounding volumes, then use gradient magnitudes from a coarse stage to decide which cameras actually matter for each block. That is a genuinely new framing for panoramic 3DGS, and the ERP-native rendering pipeline, sky-sphere modeling, and depth supervision are all well-executed. The new Pano360 dataset (5,637 images, 2M+ m²) is a real contribution — there is nothing like it for large-scale panoramic outdoor reconstruction. Code and data are promised publicly, which matters here. The ERP Jacobian derivation is clean and the experimental gains over H3DGS, CityGaussian, and DOGS are consistent across both their dataset and public benchmarks (Ricoh360, 360Roam). Model sizes are 2–7× smaller than the strongest baselines, which is a practical win. The stress-test concern about camera-per-block statistics is the real soft spot. Equation 11 assigns a camera to a block if it is geometrically inside OR if its normalized gradient score exceeds 0.8×max. In a panoramic scene, a camera could clear that threshold for many blocks simultaneously, meaning each block still gets most cameras and the “block-wise” training is barely distinguishable from global training. The paper never reports how many cameras each block actually receives relative to the total. The ablation in Table 4 shows only 0.51 dB drop when G2PS is removed entirely — surprisingly small if partitioning is truly preventing degeneration. This suggests the sky sphere, depth supervision, and ERP-native rendering are doing most of the heavy lifting, and G2PS provides marginal benefit. The τ_grad=0.8 threshold is also hand-tuned with no sensitivity analysis beyond a small table. These are fixable issues — the authors just need to report camera-per-block ratios and show the partitioning actually reduces per-block workload. The reader scored this as ACCEPT with HIGH confidence, which is about right for the method and dataset contributions. The soundness score of 7.0 is fair — the math is clean but the central scalability claim is under-evidenced. This deserves a serious referee. The dataset alone warrants review, and the method is novel enough to warrant careful checking. Recommend accept for review with a requirement that the authors substantiate the scalability claim with camera-per-block statistics.","headline":"G2PS is a solid panoramic partitioning strategy, but the paper never proves it actually reduces cameras-per-block — the scalability claim is under-evidenced.","tokens_in":13932,"tokens_out":605,"would_cite":true,"duration_ms":91729,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Gradient signals replace camera frustums for panoramic 3D scene partitioning","keywords":[],"falsifier":"If replacing frustum-based allocation with gradient-based allocation produced no improvement over naive global training (i.e., all cameras assigned to all blocks) on panoramic scenes, the core mechanism would be falsified.","tokens_in":13042,"feed_emoji":"🌐","tokens_out":1801,"duration_ms":88945,"temperature":0.7,"pith_summary":"The paper addresses a specific bottleneck in scaling 3D Gaussian Splatting to large outdoor scenes using panoramic (360°) images. Traditional partitioning strategies divide scenes into spatial blocks and assign cameras based on which cameras' viewing frustums overlap each block. Panoramic images break this: because each 360° camera sees the entire scene, every camera would be assigned to every block, collapsing block-wise optimization back into global training. The authors propose G2PS, which replaces frustum-based camera assignment with gradient-based importance scoring. After an initial coarse training pass, they measure how strongly each camera's rendering loss gradient affects the Gaussians in each spatial block. Cameras are assigned to blocks where their gradient contribution exceeds a threshold, rather than where their frustum points. Combined with parallax-driven bounding volume construction and explicit sky-sphere modeling, this enables block-wise refinement that produces models 2.2–7.5× smaller than baselines while achieving state-of-the-art rendering quality on both a new benchmark (Pano360) and two public datasets.","feed_headline":"Gradient signals replace camera frustums for panoramic 3D scene partitioning","feed_subtitle":"360° images break spatial partitioning because every camera sees everything. Gradient-based scoring restores block-wise training.","key_machinery":"Gradient-based camera-block allocation (Eq. 10–11): after coarse training, a single forward-backward pass over all training views collects per-camera, per-block gradient magnitudes. A camera is assigned to a block if either it is geometrically located within that block or its normalized gradient score exceeds a threshold (τ_grad = 0.8).","core_discovery":"The central mechanism is the substitution of gradient magnitude for geometric frustum visibility as the camera-block assignment criterion. The rendering loss gradient with respect to Gaussian positions naturally reflects how much a given viewpoint contributes to a spatial region: regions with large contributions exhibit large gradients, while distant or occluded regions produce near-zero gradients. This proxy works even when every camera technically observes every region, because gradient magnitude captures effective observational contribution rather than mere visibility. The paper shows that this single substitution, combined with a parallax-driven bounding volume and frozen sky Gaussians,恢","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Gradient magnitude replaces frustum visibility for panoramic 3DGS partitioning","Parallax-driven bounding volumes enable block-wise panoramic scene reconstruction","Loss gradients assign cameras to spatial blocks when 360° visibility breaks frustums","Geometry and gradient partitioning scales Gaussian Splatting to panoramic outdoor scenes","Gradient-based camera scoring restores block-parallel training for 360° reconstruction"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The gradient-based allocation assumes that rendering loss gradients with respect to Gaussian positions reliably measure a camera's actual contribution to a spatial block. If gradients are noisy, dominated by a few outlier Gaussians, or fail to capture true geometric importance, the camera-block assignments will be suboptimal.","fun_headline_variants_meta":{"raw":{"variants":["Gradient magnitude replaces frustum visibility for panoramic 3DGS partitioning","Parallax-driven bounding volumes enable block-wise panoramic scene reconstruction","Loss gradients assign cameras to spatial blocks when 360° visibility breaks frustums","Geometry and gradient partitioning scales Gaussian Splatting to panoramic outdoor scenes","Gradient-based camera scoring restores block-parallel training for 360° reconstruction"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":638,"prompt_tokens":544,"completion_tokens":94,"prompt_tokens_details":null},"tokens_in":544,"tokens_out":94,"duration_ms":107042,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T01:28:04.265694+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If replacing frustum-based allocation with gradient-based allocation produced no improvement over naive global training (i.e., all cameras assigned to all blocks) on panoramic scenes, the core mechanism would be falsified.","supporting_citations":[],"review_version":1}