{"id":"5bd4a4f2-fc5b-4aca-8a16-b27e3678b6a6","arxiv_id":"2507.23006","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A 3D Gaussian Splatting framework for urban scenes that combines visibility-based data partitioning, budgeted level-of-detail generation, and per-Gaussian appearance embeddings to enable efficient training and real-time rendering with robustness to appearance variation.","lead":"This paper presents a 3D Gaussian Splatting system that reconstructs large urban scenes faster and renders them in real time, even when photos have inconsistent lighting and weather. It splits the scene into regions, uses a visibility-based camera selection to cut redundant training data, and generates multiple detail levels under a strict memory budget.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison is not apples-to-apples: Supplementary D.1 lets Ours fit an appearance embedding to half of each test image before scoring the other half, while baselines get no test-image-specific adaptation.","rationale":"I read the manuscript in good faith. The method is coherent, the ablations are informative, and the code link is a plus. However, the central claim rests on quantitative comparisons whose fairness is not established. The most load-bearing point is not the private datasets or the BigCity retuning (both important), but the asymmetric test-time appearance optimization in Supplementary D.1, because it contaminates every quality comparison, including the public scenes. A single re-evaluation with a frozen appearance embedding would settle it. I therefore keep the reader's CONDITIONAL verdict: accept only after a symmetric evaluation, or after a convincing demonstration that the testing protocol is standard and its effect on the reported margins is negligible.","tokens_in":18856,"tokens_out":9053,"duration_ms":113113,"concrete_test":"On the public scenes in Table 8 (Rubble, Building, Residence, Sci-Art, Campus), re-run Ours with the appearance embedding fixed to the training-set mean and, separately, to the embedding of the nearest training camera by pose, instead of the D.1 left/right optimization. Keep all other settings identical and compare to the reported 3DGS, CityGaussian, and Hierarchical-3DGS rows. If Ours then fails to beat those baselines on a majority of scenes in PSNR or SSIM, the claimed outperformance is not established under a symmetric protocol; if the margins persist, the concern is answered.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Section 4.2) is an empirical superiority claim, and the supporting evidence is Table 1 plus Table 8. For Ours, those tables are produced under a protocol that is not applied to any baseline: Supplementary D.1 optimizes the per-image appearance embedding ℓ(I) on one half of each test image, computes metrics on the other half, and averages the two halves. Baselines (3DGS, CityGaussian, Hierarchical-3DGS, Switch-NeRF) are rendered with parameters fixed at the end of training and receive no information from the test image. Since the appearance embedding is a global latent, the left and right halves of a test image share the same illumination and weather conditions, so optimizing on one half is a real form of test-time adaptation that can inflate PSNR, SSIM, and LPIPS. The paper does not quantify this effect, and it applies to every reported quality number. The reader's concern about Supplementary C hyperparameter retuning is real but limited mainly to BigCity and partly mitigated by Table 8; this test-time asymmetry is logically prior and affects all scenes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes REUrbanGS, a 3D Gaussian Splatting framework for urban-scale scene reconstruction. It introduces point-based visibility for scene partitioning and training-camera selection, in-partition prioritized densification, a budget-controlled bottom-up LOD generation and selection scheme, an appearance transform module with per-image and per-Gaussian embeddings, and additional regularizers including depth, scale, opacity-offset, and anti-aliasing. Experiments on three large scenes (Rubble, JNU-ZH, BigCity) and four supplementary scenes compare against Switch-NeRF, 3DGS, CityGaussian, Hierarchical-3DGS, and VastGaussian, reporting PSNR, SSIM, LPIPS, number of Gaussians, and FPS. The central claim is that the method reconstructs urban scenes with better quality and efficiency than previous approaches.","tokens_in":19068,"tokens_out":6331,"duration_ms":70114,"significance":"The paper is a solid systems contribution. The visibility-based camera selection and controllable LOD generation are sensible engineering ideas, and the authors provide source code, which supports reproducibility. If the evaluation protocol were symmetric, the reported results would be evidence of a practical single-GPU pipeline for large urban scenes with real-time rendering and controllable memory use. However, the empirical claims are currently not fully supported: the quality numbers for the proposed method are obtained under a test-time adaptation protocol that baselines do not receive, baseline hyperparameters were retuned for the largest scene, and the two primary datasets are private. The significance of the contribution is therefore conditional on rectifying the evaluation.","major_comments":[{"comment":"The protocol used to compute all quality numbers for Ours in Tables 1 and 8 is asymmetric. D.1 states that the image embedding ℓ(I) is first optimized on the left half of each test image and scored on the right half, then repeated with the halves swapped. Because the two halves of the same image share illumination and weather conditions, this is genuine test-time adaptation that can inflate PSNR, SSIM, and LPIPS. No baseline receives an analogous test-image-specific adaptation, so the comparisons are not apples-to-apples. The paper's statement that this 'prevents information leakage and ensures fairness' is therefore not justified. Please report metrics with appearance embeddings fixed to a training-set value, or apply a matched test-time optimization to the baselines; at minimum, quantify the gap between the D.1 protocol and a no-adaptation protocol.","section":"Supplementary D.1"},{"comment":"For BigCity, the baselines' hyperparameters were changed substantially to fit in 80GB memory: CityGaussian's pruning ratios were raised from 40/50/60% to 70/80/90%, its densification cycle was tripled and its gradient threshold doubled; 3DGS's densification threshold was doubled; and Hierarchical-3DGS granularity was doubled from 6 to 12 pixels. These changes can plausibly degrade baseline quality beyond ordinary tuning. Table 1 shows CityGaussian LOD dropping from SSIM 0.825 (no LOD) to 0.712 on BigCity, and Hierarchical-3DGS LOD falling to 3.2 FPS. The paper should justify that these settings constitute fair tuning, or report baselines with default hyperparameters together with memory usage, or provide a sensitivity study demonstrating that the chosen settings do not materially alter the comparisons.","section":"Supplementary C"},{"comment":"The two primary large-scale datasets, JNU-ZH and BigCity, are private and not released, and no results are averaged over multiple seeds or reported with error bars. Since many reported differences are small (e.g., Rubble LPIPS 0.245 for Ours vs 0.256 for CityGaussian; FPS 99.7 vs 105.2), the absence of variance estimates and public data makes it impossible to assess whether the differences are significant or due to a favorable run. Please release the datasets or provide a detailed access mechanism, and report mean±std over at least three training runs for the headline comparisons.","section":"Section 4.1 and Tables 1, 8"},{"comment":"The abstract and Section 4.2 claim that the method 'outperforms previous approaches in both efficiency and quality,' but Table 1 does not show consistent dominance. With LOD enabled on Rubble, CityGaussian has lower #G (2.95M vs 3.60M) and higher FPS (105.2 vs 99.7); without LOD, 3DGS has lower #G and higher FPS on Rubble and JNU-ZH. The efficiency claim should be qualified (e.g., 'competitive efficiency while providing controllable budgets') or formulated as a trade-off rather than a strict improvement.","section":"Section 4.2 and Abstract"}],"minor_comments":[{"comment":"Table 3 is captioned 'Qualitative ablations' but reports quantitative metrics; please rename it to 'Quantitative ablations'.","section":"Section 4.4"},{"comment":"The caption misspells 'Hierarchial-3DGS'; it should read 'Hierarchical-3DGS'.","section":"Figure 5 caption"},{"comment":"The phrase 'the the Fundamental Research Funds' contains a duplicated article and should be corrected.","section":"Acknowledgments"},{"comment":"The notation 'knn_{i;k}' in Equation (9) is not defined; please state explicitly that k is the number of nearest neighbors and clarify how the normalization over pairs is computed.","section":"Section 3.5.1"},{"comment":"The header 'Budget (×100)' is ambiguous; please state whether the listed values are already multiplied by 100 or are in units of 100, and give the actual budget values used.","section":"Table 2"},{"comment":"The entry 'N/A (encountered a bug)' for CityGaussian on Campus should be replaced with a description of the failure mode or removed, since it is not a quantitative result.","section":"Supplementary E.2, Table 8"}],"recommendation":"major_revision","confidential_remarks":"The test-time adaptation asymmetry in Supplementary D.1 is the most serious issue and should be addressed before publication. The method itself appears well-engineered and the issue is fixable by reporting no-adaptation metrics or applying a matched protocol to baselines; releasing the private datasets would further strengthen the paper. I do not see a correctness error that would force rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid engineering paper that assembles known 3DGS scaling techniques into one urban reconstruction pipeline. The genuinely new pieces are modest — convex-hull visibility for image-to-partition assignment, distance-aware densification thresholds, bottom-up budgeted LOD generation, and per-Gaussian plus per-image appearance embeddings. The paper is straightforward about what it borrows (Block-NeRF, Taming3DGS, CityGaussian, Mip-Splatting, AbsGS, DNGaussian, NeRF-W), and the supplement gives enough detail on hyperparameters for someone to reproduce the pipeline. If the code delivers what the project page promises, this will be a useful reference implementation for urban-scale Gaussian splatting.\n\nThe soft spots are real, and one is more fundamental than the reader's report suggests. Supplementary D.1 describes a test-time protocol where the appearance embedding is optimized on one half of each test image and then scored on the other half, with the two halves averaged. No baseline gets any test-image-specific adaptation. Since appearance embeddings capture global illumination and weather, optimizing on one half of the test image is genuine test-time adaptation, and it can inflate PSNR, SSIM, and LPIPS. This asymmetry is logically prior to the BigCity hyperparameter concern: it affects every reported quality number, not just the largest scene. The paper does not quantify how much the numbers change without this protocol, and it applies to the Ours rows in every table.\n\nThe BigCity baseline retuning in Supplementary C is also a legitimate worry, though slightly narrower: doubling pruning ratios for CityGaussian and doubling granularity for Hierarchical-3DGS to fit in 80GB could unfairly handicap those baselines. Table 8 on public scenes partially mitigates the worry, but the two main datasets are private, and no error bars or multiple seeds are reported. The method itself is coherent, and the ablations do show each component matters, so I don't think the paper is wrong in its architecture. I just don't think the headline \"outperforms previous approaches in both efficiency and quality\" is supported as written.\n\nWho this is for: anyone building deployable large-scale 3DGS systems, especially for autonomous driving or city digital twins, will get a practical checklist of components and a plausible default pipeline. It deserves a serious referee: the engineering is sound, the supplement is honest, and the evaluation flaw is fixable. I would send it to review, with a request to rerun the comparisons without test-time embedding optimization, add error bars, and ideally validate on a public dataset. If the numbers hold after that, it becomes a solid systems contribution.","headline":"A competent engineering pipeline for urban 3DGS whose headline comparison is undermined by a test-time appearance-embedding adaptation that favors the proposed method on every reported metric.","tokens_in":19633,"tokens_out":2616,"would_cite":true,"duration_ms":31305,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that city-scale 3D scenes can be reconstructed from drone images on a single 24 GB GPU, rendered in real time, with higher quality than prior 3D Gaussian Splatting methods.","keywords":["3D Gaussian Splatting","urban scene reconstruction","level-of-detail rendering","appearance transform","visibility-based image selection","resource-constrained training","real-time rendering","drone photogrammetry"],"falsifier":"Run the reported benchmark on a public release of one of the two private drone scenes, or on another public urban scene of comparable scale, with each baseline tuned by its original authors under a 24 GB memory cap; if the reported PSNR, SSIM, LPIPS, and FPS gaps shrink or reverse, the claim fails. A cheaper check is to retrain the two LOD baselines on BigCity with their default pruning and granularity settings to see whether the Supplementary C hyperparameter adjustments, rather than the method itself, produced the quality gap.","tokens_in":18628,"feed_emoji":"🏙️","tokens_out":12002,"duration_ms":127758,"temperature":0.7,"pith_summary":"The paper tries to establish that 3D Gaussian Splatting, previously practical mainly for small scenes, can be turned into a city-scale reconstruction system that runs on a single 24 GB GPU and still renders at real-time rates. It claims this is achieved by making every stage resource-aware: images are assigned to partitions by point-based visibility rather than location alone, densification is concentrated inside each partition with a distance-dependent gradient threshold, and a bottom-up level-of-detail generator builds each detail level under a strict Gaussian budget. A per-Gaussian appearance transform module absorbs lighting, season, and camera differences that normally make 3DGS grow floating artifacts, while depth, scale, and anti-aliasing regularizers suppress floaters and jagged edges. The reported experiments on three aerial scenes, two of them private drone captures, show higher PSNR, SSIM, and LPIPS than four prior methods while using modest Gaussian counts and real-time FPS. If the claim holds, the method offers a practical single-GPU recipe for arbitrarily large urban reconstruction with user-controllable resource use.","feed_headline":"One 24 GB GPU rebuilds city-scale scenes in real time","feed_subtitle":"Controllable levels of detail and per-image appearance fixes beat prior methods on quality and speed.","key_machinery":"Two mechanisms carry the efficiency claim and one carries the robustness claim. The efficiency core is the controllable bottom-up level-of-detail generator: a sequence of budgets $B_1<B_2<\\cdots<B_l$, densification intervals $T_1>T_2>\\cdots>T_l$, and downsampling factors $D_1<D_2<\\cdots<D_l=1$; training starts at the coarsest level and checkpoints upward, so each level refines the previous one while the Gaussian count never exceeds the user-set budget for that level. Supporting it are point-based visibility, the fraction of the projected SfM point cloud's convex hull that an image sees within a partition, which selects training cameras and rebalances partitions, and in-partition prioritized densification, which allows a Gaussian at distance $d_i$ from its partition to densify only when its gradient exceeds $\\hat{\\tau}_{\\min}(\\frac{\\min(d_i,\\hat{d}_{\\max})}{\\hat{d}_{\\max}}(\\eta-1)+1)$, concentrating computation inside the partition. The robustness core is the appearance transform module: learned image embeddings and Gaussian embeddings passed through a lightweight MLP that predicts per-Gaussian color and opacity offsets, so appearance variations are modeled explicitly instead of being baked into extra Gaussians.","core_discovery":"The central claim is that an urban-scale scene can be reconstructed from multi-view drone imagery without post-training compression, in less time and at higher fidelity than previous 3D-Gaussian approaches, by replacing uncontrolled densification with a budgeted level-of-detail hierarchy. The paper shows that three resource controls are sufficient: visibility-based camera assignment removes redundant training views; in-partition prioritized densification directs Gaussians to where they matter; and bottom-up LOD generation produces a sequence of levels, each constrained by a budget $B_i$, densification interval $T_i$, and image downsampling factor $D_i$, so that rendering can switch levels by partition-camera distance. On top of this, the appearance transform module learns a per-image and per-Gaussian embedding, feeding a lightweight MLP that predicts color and opacity offsets, which removes inter-image appearance inconsistencies and permits post-reconstruction appearance changes at no rendering cost. The experiments report that the LOD-enabled model renders the three test scenes at 99.7, 63.9, and 73.0 FPS with 3.60M, 6.65M, and 6.84M Gaussians while improving quality metrics over the compared LOD systems.","pith_inferences":["A natural check not performed in the paper is to release the two private scenes and run the same protocol on a public urban dataset; if the gains survive, the method becomes a reproducible benchmark rather than a private-data demonstration.","The per-Gaussian appearance embeddings point toward editable digital twins: by swapping only the image embedding, the same reconstruction could display different seasons, lighting, or camera styles without re-rendering cost, an ability the paper demonstrates for two building states.","The visibility-based camera selection is a generic data-reduction idea that could be transplanted to other partition-based scene representations, such as grid- or hash-based NeRF variants, to cut their training data as well.","Because the LOD levels are checkpoints with increasing budgets, they could be streamed progressively to a viewer—coarse level first, finer levels on approach—which would suit low-bandwidth city-scale browsing; the paper itself notes its current LOD switching is not incremental, so this is an extension rather than a claim."],"forward_implications":["LOD mode gives a direct user-controlled trade-off between fidelity and footprint: cutting the budget on Rubble from $(8192,16384,32768)$ to $(1024,2048,4096)$ reduces Gaussians from 3.80M to 1.61M and raises FPS from 96.4 to 126.9 while PSNR drops only from 27.11 to 26.13.","Visibility-based camera selection removes 1.28–1.70 times as many redundant training views as location-only assignment, so the same quality is reached with fewer training iterations on the tested scenes.","In-partition prioritized densification accelerates training by 1.30–1.64 times and cuts the final Gaussian count by 1.37–2.05 times, with negligible effect on quality metrics.","The appearance transform module improves all three quality metrics on all three scenes and typically reduces Gaussian count, because it prevents the model from growing floaters to overfit appearance variation.","Except for the non-LOD configuration, all training and rendering steps in the reported experiments run within 24 GB of GPU memory, so the pipeline is usable on a single commodity card."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian representation, rasterization, and loss that the whole pipeline starts from.","marker":"[8]"},{"why":"Provides the steerable densification with resource budgets on which the controllable LOD generation is built.","marker":"[20]"},{"why":"Supplies the divide-and-conquer scene-partitioning layout that the paper extends with visibility-based camera selection and rebalancing.","marker":"[41]"},{"why":"Provides the partition-level LOD selection scheme the rendering stage adopts, and is the main large-scale LOD baseline.","marker":"[18]"},{"why":"Provides the hierarchical Gaussian representation and LOD baseline for very large datasets that the method is compared against.","marker":"[9]"},{"why":"Provides the Rubble and Building test scenes, camera poses, and SfM pipeline used for evaluation.","marker":"[42]"},{"why":"Provides the Residence, Sci-Art, and Campus public urban test scenes used in supplementary comparisons.","marker":"[17]"},{"why":"Supplies the per-image appearance embedding idea that the appearance transform module refines to per-Gaussian embeddings.","marker":"[21]"}],"fun_headline_variants":["LOD-controlled Gaussian splatting rebuilds cities in real time","Real-time urban 3D from multi-view captures on one GPU","Fast city-scale reconstruction with appearance-robust Gaussian splatting","Budgeted Gaussian density yields fast, robust urban scene 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The superiority claim rests on the fairness of the baseline hyperparameter adjustments made in Supplementary C and on the public availability of the two largest test scenes; if the adjusted baselines were handicapped more than fair tuning would, or if the private datasets hide unfavorable conditions, the measured gap could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["LOD-controlled Gaussian splatting rebuilds cities in real time","Real-time urban 3D from multi-view captures on one GPU","Fast city-scale reconstruction with appearance-robust Gaussian splatting","Budgeted Gaussian density yields fast, robust urban scene 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3433,"prompt_tokens":927,"completion_tokens":2506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2431}},"tokens_in":543,"tokens_out":2506,"duration_ms":20693,"temperature":1.0,"reasoning_tokens":2431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:08:33.363598+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reported benchmark on a public release of one of the two private drone scenes, or on another public urban scene of comparable scale, with each baseline tuned by its original authors under a 24 GB memory cap; if the reported PSNR, SSIM, LPIPS, and FPS gaps shrink or reverse, the claim fails. A cheaper check is to retrain the two LOD baselines on BigCity with their default pruning and granularity settings to see whether the Supplementary C hyperparameter adjustments, rather than the method itself, produced the quality gap.","supporting_citations":[{"cited_title":"Taming 3dgs: High-quality radiance fields with limited resources","cited_arxiv_id":null,"evidence_quote":"Provides the steerable densification with resource budgets on which the controllable LOD generation is built."},{"cited_title":"Block-nerf: Scalable large scene neural view synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the divide-and-conquer scene-partitioning layout that the paper extends with visibility-based camera selection and rebalancing."},{"cited_title":"Citygaussian: Real-time high-quality large-scale scene rendering with gaussians","cited_arxiv_id":null,"evidence_quote":"Provides the partition-level LOD selection scheme the rendering stage adopts, and is the main large-scale LOD baseline."},{"cited_title":"A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the hierarchical Gaussian representation and LOD baseline for very large datasets that the method is compared against."},{"cited_title":"Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs","cited_arxiv_id":null,"evidence_quote":"Provides the Rubble and Building test scenes, camera poses, and SfM pipeline used for evaluation."},{"cited_title":"Capturing, reconstructing, and simulating: the urbanscene3d dataset","cited_arxiv_id":null,"evidence_quote":"Provides the Residence, Sci-Art, and Campus public urban test scenes used in supplementary comparisons."},{"cited_title":"Nerf in the wild: Neural radiance fields for uncon- strained photo collections","cited_arxiv_id":null,"evidence_quote":"Supplies the per-image appearance embedding idea that the appearance transform module refines to per-Gaussian embeddings."}],"review_version":1}