{"id":"400bda92-d364-477f-9c4a-d15fd94a4b96","arxiv_id":"2505.21041","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid of proxy building meshes and residual plus surrounding 3D Gaussians renders large urban scenes in real time on mobile GPUs with substantially less memory than pure 3DGS.","lead":"CityGo combines simplified building meshes with two kinds of 3D Gaussian splats to render city-scale aerial scenes in real time on mobile GPUs. The result is a modeling pipeline that trains about 1.4 times faster than pure 3D Gaussian Splatting while using a fraction of the memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1.4x speedup and 'comparable fidelity' claims are confounded by unmatched SH order (CityGo SH=0 vs 3DGS SH=2); without an SH-matched baseline the speedup may be attributed to SH reduction rather than the hybrid representation.","rationale":"The reader's verdict correctly identifies that the SH mismatch weakens the fidelity comparison, but the 'weakest assumption' field points to proxy geometry and GIS mask accuracy. The proxy-geometry limitation is acknowledged by the authors in the conclusion ('Non-building structures... may be misclassified... resulting in artifacts and PSNR degradation'), and it bears on robustness rather than on the validity of the headline numbers under the tested conditions. In contrast, the SH-order mismatch directly affects the experimental support for both the speedup and the 'comparable fidelity' claim. If the speedup disappears when the baseline is SH-matched, the central claim of the paper is substantially weakened. This is not a stylistic preference: the paper itself states that SH=0 is used to accelerate convergence, making it a genuine experimental confound. A single ablation of SH order would settle the issue, so the correct verdict remains CONDITIONAL rather than ACCEPT or REJECT. I do not see internal inconsistency in the method itself; the concern is about the fairness of the comparison. The reader's rationale mentions the SH confound, so my agreement is partial rather than total.","tokens_in":13342,"tokens_out":8650,"duration_ms":100157,"concrete_test":"Re-run Table 2 with two extra conditions on Area-H and Area-L: (a) pure 3DGS with SH=0, same 100K iterations, same block overlap and merging; (b) CityGo with SH=2 for the initial, residual, and surrounding Gaussians, keeping all other settings identical. If condition (a) reaches a training time close to CityGo's, or condition (b) loses most of the speedup relative to 3DGS-SH2, then the claimed 1.4x speedup is primarily an SH-order effect, not a benefit of the hybrid representation. Report PSNR, training time, and FPS for both conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a 1.4x training speedup with comparable visual fidelity to pure 3DGS is directly undermined by an experimental control gap: in Section 5.2, 3DGS and 2DGS are trained with SH=2, while CityGo's Gaussians use SH=0. The paper explicitly uses SH=0 'to accelerate convergence' (Section 5.1), so the SH choice is a known speed lever. Lower SH reduces per-Gaussian parameters and training cost, and removes view-dependent appearance, making the speed comparison and the 'comparable fidelity' claim systematically favorable to CityGo. The paper does not report any SH-matched ablation, such as 3DGS with SH=0 or CityGo with SH=2. Section 4.1 of the reader's verdict flags the same confound but does not make it the weakest assumption. Without this control, the headline speedup cannot be attributed to the hybrid proxy-plus-residual representation rather than to a cheaper appearance model. This is an internal controls issue, not a disagreement with any external consensus, and it is testable with a small set of additional runs.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CityGo proposes a hybrid representation for large-scale aerial urban scenes that combines textured proxy meshes for buildings with two types of 3D Gaussians: residual Gaussians that capture high-frequency details on building surfaces and downsampled surrounding Gaussians for the rest of the environment. The pipeline starts from MVS point clouds, completes building geometry via a layer-based point-cloud completion method (BPCC), extracts planar proxy meshes via FPP/KSR, generates occlusion-free textures using zero-order SH Gaussians and TwinTex-style rendering, and then jointly optimizes proxy textures and Gaussian parameters. The paper reports a roughly 1.4x training-time speedup over pure 3DGS, comparable PSNR, large reductions in model size and Gaussian count, and real-time rendering on a Jetson AGX Orin.","tokens_in":13598,"tokens_out":4110,"duration_ms":45476,"significance":"If the reported efficiency and fidelity results hold, CityGo would be a practically useful step toward city-scale photorealistic rendering on edge devices, and the hybrid mesh-plus-residual-Gaussian representation is a sensible design direction for aerial urban data. The paper has concrete strengths: it evaluates on two large drone-captured scenes (1.5 km² each) plus the UrbanBIS benchmark, reports mobile GPU FPS, and quantifies the proxy-mesh compression (61.7x fewer vertices/faces). The ablation on UV finetuning and CRM-based sampling is useful, and the hybrid rendering formulation in Eq. (4) is explicit. However, the central quantitative claims rest on comparisons in which the SH order is not matched between CityGo and the Gaussian baselines, which makes the headline speedup and 'comparable fidelity' statements not yet established.","major_comments":[{"comment":"The main comparison is confounded by an unmatched SH order. The paper states in Section 5.1 that CityGo uses SH=0 'to accelerate convergence,' while Section 5.2 reports that the 3DGS and 2DGS baselines are trained with SH=2. SH=0 removes view-dependent appearance and reduces the per-Gaussian parameter count and per-pixel evaluation cost, which directly lowers both training time and memory. The 1.14-1.4x training speedup and the 'comparable fidelity' claim can therefore be attributed partly or entirely to the cheaper appearance model rather than to the hybrid proxy-plus-residual representation. This is an internal control gap, and it is testable. I strongly recommend adding an SH-matched ablation: either 3DGS with SH=0 (and 2DGS with SH=0) or a CityGo variant with SH=2, reporting PSNR, training time, model size, and FPS. Without this, the central claim of a 1.4x speedup from the hybrid representation is not supported.","section":"Section 5.2, Tables 1-2, Section 5.1"},{"comment":"It is unclear what is included in the reported training time for CityGo. The pipeline described in Sections 3-4 has multiple stages: initial SH=0 3DGS training, BPCC point-cloud completion, FPP/KSR mesh extraction, synthetic texture rendering via TwinTex, UV finetuning, residual Gaussian selection and 100K-iteration optimization, and surrounding Gaussian downsampling with 30K-iteration optimization. If 'Time' for CityGo counts only the Gaussian optimization stages, while the time for 3DGS counts the full end-to-end training, the speedup comparison is not apples-to-apples. Please specify the timing methodology, list the per-stage timings, and state whether mesh extraction and texturing are included in the reported 17.6 h / 16.6 h / 1.4 h figures.","section":"Section 5.2, 'Time (h)' columns"},{"comment":"The abstract's claim of 'on average 1.4x speedup' is not fully supported by the tables. The speedups implied by the reported training times are 1.6/1.4 = 1.14x on UrbanBIS (Table 1), 24.3/17.6 = 1.38x on Area-H, and 23.4/16.6 = 1.41x on Area-L (Table 2). The average across all three datasets is approximately 1.31x, not 1.4x. If the '1.4x' is intended to describe only the two aerial datasets, that should be stated explicitly; otherwise the abstract's quantitative claim should be corrected to match the reported data.","section":"Abstract, Tables 1-2"}],"minor_comments":[{"comment":"The BPCC thresholds gamma (area ratio for introducing a new dominant contour) and beta (hole-filling ratio) are described as fixed, but no values are given and no sensitivity analysis is provided; please report the chosen values and, ideally, an ablation or robustness check.","section":"Section 3.1"},{"comment":"In Eq. (5), the notation P_gamma for the set of rays from 'viewpoint-gamma' is inconsistent with the later use of gamma as a threshold; please rename one of these to avoid confusion.","section":"Section 3.3, Eq. (5)"},{"comment":"The FPS numbers in Table 2 are measured at 1988x1326 on an A6000, while Table 3 uses 720p on a Jetson AGX Orin; the caption of Table 3 states the resolution, but the main text does not draw attention to this resolution difference when discussing real-time performance.","section":"Section 5.2"},{"comment":"The ablation table reports PSNR and size but no error bars or multiple runs; although the differences are small, stating how many runs were averaged (if any) would strengthen the ablation.","section":"Section 5.3, Table 4"},{"comment":"There are several typos and grammatical slips, e.g., 'UA Vs' in the Introduction, 'comsumption' in Section 5.2, 'based using CRM' in Section 3.3, and inconsistent capitalization 'CityGO' in the captions of Figures 7 and 8.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The SH-order confound is the main technical obstacle; it is a straightforward control experiment and, if the authors provide SH-matched baselines, the paper could become acceptable. The timing-methodology ambiguity is also important and should be fixed with a clear breakdown. I would not reject on current evidence, but the central speedup/fidelity claims are not yet verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: CityGo is a genuine engineering contribution that clearly reduces memory and achieves real-time rendering on edge hardware, but the headline 1.4x training-time speedup over pure 3DGS is not clean because CityGo uses zero-order SH while the 3DGS baseline uses SH=2. That is an experimental control gap, not a fatal flaw.\n\nWhat is actually new: the assembly of textured proxy meshes, residual Gaussians placed by color residual maps, and importance-downsampled surrounding Gaussians is a new combination for aerial urban scenes. The BPCC modification to [20] with bottom-up clustering and alpha-shape contours is a sensible fix for concave buildings. The guard-interval depth constraint that ties Gaussian placement to proxy geometry is a reasonable way to keep the hybrid stable. And the results show real wins: memory footprint drops to roughly 1/8 of 3DGS, FPS on a Jetson is 20-51, and the training-time reduction is consistent across three datasets (1.14-1.4x). That consistency suggests the method itself is not a fluke.\n\nThe soft spots are in the evaluation. The SH mismatch is the big one: lower SH cuts per-Gaussian storage and training cost and removes view-dependent effects, which makes both the speedup and the 'comparable fidelity' claim systematically favor CityGo. There is no 3DGS-with-SH=0 or CityGo-with-SH=2 ablation, so you cannot attribute the speedup to the hybrid representation rather than the cheaper appearance model. Also missing: error bars or multiple seeds (these are single runs), and no code or data release, which limits independent verification. The authors do honestly flag the reliance on accurate proxy geometry and GIS masks, and the fixed-opacity texture issue, but those are real limitations, especially for scenes with cranes or signage.\n\nThe paper is written clearly and the pipeline is coherent. The central idea holds up qualitatively; the quantitative head-to-head needs a matched baseline. This is exactly the kind of work a serious referee can improve with a few additional experiments. I would send it to peer review and ask for the SH-matched control, error bars, and code release.\n\nRegards,","headline":"A promising hybrid mesh-Gaussian urban renderer with real efficiency gains, but the central speedup claim is confounded by an unmatched SH order and needs a control run.","tokens_in":14173,"tokens_out":2530,"would_cite":true,"duration_ms":28377,"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":"CityGo combines textured proxy buildings with residual and surrounding 3D Gaussians to make large-scale aerial urban scenes render in real time on mobile GPUs, cutting training time by about 1.4x compared with pure 3D Gaussian Splatting…","keywords":["hybrid scene representation","3D Gaussian Splatting","proxy geometry","urban modeling","aerial imagery","real-time rendering","mobile GPU","texture mapping"],"falsifier":"Run CityGo on a dense urban scene containing a large non-building structure attached to a building (e.g., a tower crane or a large sign). If the segmentation mistakenly treats that structure as part of the building, the fixed-opacity proxy texture will bake its color onto the facade, and held-out-view PSNR should drop by substantially more than the roughly 1 dB gap CityGo already accepts versus 3DGS—an observation that would confirm the stated limitation and show that the hybrid representation loses its fidelity advantage exactly when proxy geometry fails.","tokens_in":13107,"feed_emoji":"🏙️","tokens_out":3278,"duration_ms":35059,"temperature":0.7,"pith_summary":"CityGo claims that large-scale urban scenes from aerial imagery can be modeled far more cheaply than pure 3D Gaussian Splatting allows, by splitting the scene into two parts: buildings become lightweight textured proxy meshes, while everything else stays as sparse 3D Gaussians. The paper shows that this hybrid representation trains about 1.4 times faster than standard 3DGS, uses dramatically less memory, and still renders 1.5 km² scenes at a minimum of 20 FPS on a Jetson AGX Orin mobile GPU. If true, this makes photorealistic city-scale digital twins, AR navigation, and UAV inspection feasible on edge devices rather than only on high-end desktop GPUs.","feed_headline":"Hybrid city renderer trains 1.4x faster, runs real-time on mobile GPU","feed_subtitle":"Textured proxy buildings plus residual Gaussians render 1.5-km² scenes at 20 FPS on an integrated GPU with near-3DGS quality.","key_machinery":"The load-bearing mechanism is the three-part hybrid representation. Building proxy meshes are extracted from MVS point clouds through a bottom-up, cluster-aware version of layer-based proxy reconstruction (Building Point Cloud Completion), then refined by planar primitive fitting and kinetic shape reconstruction; textures are generated from zero-order SH Gaussian renderings via the TwinTex method and refined with differentiable UV optimization. Residual Gaussians are selected by a color residual map score, which measures where proxy-rendered color disagrees with ground truth photos, and their placement is constrained near the mesh surface using a depth guard interval. Surrounding Gaussians are kept from the original full-scene Gaussians but importance-sampled (following Mini-Splatting) to remove redundancy. Together, these components make the representation both compact enough for mobile inference and fine-grained enough to recover facade details.","core_discovery":"The central claim is that a hybrid scene representation—compact textured proxy geometry for buildings, sparse residual Gaussians for high-frequency building details, and importance-downsampled surrounding Gaussians for unstructured environment—recovers the visual quality of dense 3D Gaussian Splatting at a fraction of the cost. The authors demonstrate on aerial datasets and the UrbanBIS benchmark that their method achieves a 1.4x average training speedup, reduces model size to roughly 1/8 of 3DGS on their own captures, and sustains real-time rendering (161–196 FPS on an RTX A6000, 20–51 FPS on a mobile GPU) with PSNR within about 1 dB of 3DGS. The key insight is that buildings, which dominate city scenes, can be represented as stable planar-proxy meshes with baked textures, leaving only the subtle, view-dependent, and geometrically complex parts to be carried by Gaussians.","pith_inferences":["The same hybrid logic could extend beyond buildings: any class with predictable planar structure (e.g., roads, bridges, parking structures) could be absorbed into proxy geometry, potentially shrinking the Gaussian budget even further on infrastructure-heavy scenes.","Because the color residual map explicitly exposes where the proxy fails, CityGo's component selection is effectively an automatic budget allocator: it spends Gaussian capacity exactly where the mesh cannot reproduce the photograph, which suggests it could be tuned against a perceptual loss rather than PSNR to yield even smaller footprints.","A testable extension would be to replace the fixed proxy opacity of 1.0 with an optimizable transparency, which would let the optimization stage correct misclassified non-building structures instead of baking their colors into the facade—an evolution the paper's own limitation section points toward."],"forward_implications":["If CityGo's claims hold, city-scale digital twins and AR navigation layers can be deployed on smartphones, drones, and in-vehicle hardware without a cloud backend, since rendering is 20+ FPS on a mobile GPU and training runs on a single desktop GPU.","The 1.4x training speedup and roughly 1/8 model-size reduction relative to 3DGS mean that urban scenes can be reconstructed and updated frequently, making practical the kind of 'living city' models that would otherwise require costly re-optimization.","The hybrid representation decouples structured buildings, which can be edited or replaced, from unstructured surroundings, so urban-planning tools could swap in new building designs while keeping the Gaussian environment intact."],"supporting_citations":[{"why":"Defines the baseline 3D Gaussian Splatting representation, training procedure, and rendering pipeline that CityGo builds on and compares against.","marker":"[23]"},{"why":"Supplies the off-the-shelf SfM/MVS pipeline that produces the dense point clouds used to initialize both the baseline Gaussians and the building proxy extraction.","marker":"[1]"},{"why":"Provides the layer-based proxy geometry reconstruction approach that BPCC adapts and improves with dense input, clustering, and alpha-shape contours.","marker":"[20]"},{"why":"The TwinTex method used to generate occlusion-free building textures from the Gaussian-rendered synthetic views.","marker":"[45]"},{"why":"The Fitting Planar Primitives algorithm that approximates completed building point clouds with planar surfaces before kinetic shape reconstruction.","marker":"[49]"},{"why":"Kinetic Shape Reconstruction, which partitions space into inside/outside regions and produces the final proxy building mesh geometry.","marker":"[2]"},{"why":"Mini-Splatting's importance-based sampling strategy is adopted to downsample surrounding Gaussians while preserving perceptual quality.","marker":"[12]"},{"why":"CityGaussianV2 serves as a large-scale urban rendering baseline whose quality, size, and speed are directly compared in the experiments.","marker":"[30]"}],"fun_headline_variants":["Hybrid proxy-Gaussian city renderer trains 1.4x faster, uses 1/8 memory","CityGo: real-time urban rendering on mobile with proxy buildings and residual Gaussians","1.4x faster, 8x less memory: hybrid city rendering for mobile","Proxy buildings plus sparse Gaussians: city-scale rendering on mobile GPUs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that the proxy-building extraction—especially the GIS building masks and the MVS point cloud completion—produces geometry accurate enough that the fixed-opacity baked textures do not need to be corrected by the residual Gaussians.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid proxy-Gaussian city renderer trains 1.4x faster, uses 1/8 memory","CityGo: real-time urban rendering on mobile with proxy buildings and residual Gaussians","1.4x faster, 8x less memory: hybrid city rendering for mobile","Proxy buildings plus sparse Gaussians: city-scale rendering on mobile GPUs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3496,"prompt_tokens":1066,"completion_tokens":2430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":2338}},"tokens_in":682,"tokens_out":2430,"duration_ms":16369,"temperature":1.0,"reasoning_tokens":2338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:38:15.619826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CityGo on a dense urban scene containing a large non-building structure attached to a building (e.g., a tower crane or a large sign). If the segmentation mistakenly treats that structure as part of the building, the fixed-opacity proxy texture will bake its color onto the facade, and held-out-view PSNR should drop by substantially more than the roughly 1 dB gap CityGo already accepts versus 3DGS—an observation that would confirm the stated limitation and show that the hybrid representation loses its fidelity advantage exactly when proxy geometry fails.","supporting_citations":[{"cited_title":"3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Trans","cited_arxiv_id":null,"evidence_quote":"Defines the baseline 3D Gaussian Splatting representation, training procedure, and rendering pipeline that CityGo builds on and compares against."},{"cited_title":"Agisoft photoscan user manual: Professional edi- tion.https://www.agisoft.com, 2016","cited_arxiv_id":null,"evidence_quote":"Supplies the off-the-shelf SfM/MVS pipeline that produces the dense point clouds used to initialize both the baseline Gaussians and the building proxy extraction."},{"cited_title":"Fast Building Instance Proxy Reconstruction for Large Urban Scenes.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the layer-based proxy geometry reconstruction approach that BPCC adapts and improves with dense input, clustering, and alpha-shape contours."},{"cited_title":"TwinTex: Geometry-Aware Texture Generation for Abstracted 3D Ar- chitectural Models.ACM Trans","cited_arxiv_id":null,"evidence_quote":"The TwinTex method used to generate occlusion-free building textures from the Gaussian-rendered synthetic views."},{"cited_title":"Finding good configurations of planar primitives in unorganized point clouds","cited_arxiv_id":null,"evidence_quote":"The Fitting Planar Primitives algorithm that approximates completed building point clouds with planar surfaces before kinetic shape reconstruction."},{"cited_title":"Kinetic Shape Reconstruction.ACM Trans","cited_arxiv_id":null,"evidence_quote":"Kinetic Shape Reconstruction, which partitions space into inside/outside regions and produces the final proxy building mesh geometry."},{"cited_title":"Mini-Splatting: Represent- ing Scenes with a Constrained Number of Gaussians","cited_arxiv_id":null,"evidence_quote":"Mini-Splatting's importance-based sampling strategy is adopted to downsample surrounding Gaussians while preserving perceptual quality."},{"cited_title":"CityGaussianV2: Efficient and Geomet- rically Accurate Reconstruction for Large-Scale Scenes","cited_arxiv_id":null,"evidence_quote":"CityGaussianV2 serves as a large-scale urban rendering baseline whose quality, size, and speed are directly compared in the experiments."}],"review_version":1}