{"id":"19b0a9ba-0fee-4bc6-909f-3510fd7f76ee","arxiv_id":"2607.03771","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Viewpoint-orientation partitioning of cameras plus selective completion of sparse SfM points enables higher-quality large-scale surface meshes from 3DGS than spatial-block baselines.","lead":"This paper introduces viewpoint-orientation clustering plus targeted point-cloud hole filling to make 3D Gaussian Splatting produce accurate city-scale surfaces. The approach yields higher F1 scores than recent block-based 3DGS methods on three public large-scene benchmarks while remaining multi-GPU friendly.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged threshold sensitivity.","rationale":"The paper's strongest claim is an empirical engineering result, not a theoretical necessity. The orientation-first clustering (cosine DBSCAN then position K-means) is a clear, well-motivated departure from pure spatial blocks, and the multi-view photometric/geometric losses inside each group are standard and correctly applied. Ablations already bound the sensitivity of the two hand-tuned thresholds; residual gains persist. Protocol issues (cropped evaluation, no code, no significance tests) justify the reader's CONDITIONAL verdict but do not introduce a new load-bearing attack that would move the verdict to REJECT or require a different condition. Therefore the stress-test leaves the reader's assessment intact.","tokens_in":15024,"tokens_out":523,"duration_ms":4514,"concrete_test":"Re-run the full GauU-Scene pipeline with ε_dir swept over {0.5, 0.71, 0.85} and τ over {0.6, 0.7, 0.8} on the uncropped GT point clouds (if released); if mean F1 remains ≥ CityGS-X under at least two of the three ε_dir values, the threshold-sensitivity concern does not overturn the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (orientation partitioning + selective completion yields higher F1 than CityGaussian-V2 / CityGS-X) is supported by the reported numbers and ablations. The reader's weakest assumption (fixed ε_dir=0.71 / τ=0.7 may be scene-dependent) is real but already quantified: Table 4 shows F1 drops only to 0.532–0.544 outside 45°, still above CityGS-X's 0.527; Table 5 and the w/o-completion row show the method remains competitive even without hole filling. No deeper internal inconsistency (e.g., circular evaluation, unstated data leakage, or contradiction between partitioning logic and multi-view losses) is evident in Sections 3.1–3.3 or the tables. Cropped GT regions and missing code are protocol limitations already noted by the reader, not load-bearing flaws in the argument itself.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes VOP-GS for city-level multi-view surface reconstruction with 3D Gaussian Splatting. It partitions input views by camera orientation (DBSCAN on cosine distance of viewing directions, Eq. 1, with ε_dir=0.71) followed by position-based K-means splitting when clusters exceed M_max=200, then refines groups by adding favorable source views. It further detects holes in the COLMAP sparse point cloud via normalized Shannon entropy of projected keypoints on an N\times N grid (Eqs. 3–4, τ=0.7) and completes them by triangulating dense matches from a pretrained network (MINIMA). Each group is optimized independently with multi-view photometric/geometric losses from PGSR plus standard RGB and depth-normal terms (Eq. 5). Experiments on GauU-Scene, MatrixCity, and UrbanScene3D claim higher F1 surface accuracy than CityGaussian-V2 and CityGS-X, with competitive multi-GPU training time.","tokens_in":15262,"tokens_out":1309,"duration_ms":17591,"significance":"If the reported F1 gains hold under fair evaluation, the work provides a simple, geometry-oriented alternative to purely spatial block partitioning for large-scale 3DGS surface reconstruction. Orientation clustering directly targets the multi-view overlap that prior surface methods have shown improves depth estimation, while selective (not global) point-cloud completion avoids the cost of full MVS. Consistent quantitative improvements across three datasets, supported by ablations on angular threshold (Table 4) and entropy τ (Table 5), plus efficiency numbers (Table 3), make the approach practically useful for multi-GPU city-scale pipelines. The method is largely engineering but addresses a genuine gap: most large-scale 3DGS systems optimize for novel-view synthesis rather than metric surfaces.","major_comments":[{"comment":"§4.4 and Table 1: Quantitative claims rest on cropped central regions of the GauU-Scene LiDAR GT (following CityGaussian-V2) because edge regions are incomplete. This protocol is not neutral; completion of the initialization point cloud can occur outside the retained crop, and the paper itself notes that CityGaussian-V2 can look more complete yet score lower. Report uncropped F1 (or a fixed, publicly documented crop mask) so that completeness and accuracy can be assessed independently.","section":"§4.4 / Table 1"},{"comment":"§4.4 and Table 2: The CityGS-X numbers obtained by the authors (F1 0.527 on GauU-Scene, 0.581 on MatrixCity) diverge from those in the original CityGS-X paper, and the authors remark that the meshes “do not match those shown in its original paper.” Without a verified re-implementation or shared evaluation script, the SOTA ranking is not fully reliable. Either release the exact evaluation code/masks used for all baselines or re-run CityGS-X under identical conditions and document any hyper-parameter differences.","section":"§4.4 / Table 2"},{"comment":"§3.1 (Sub-cluster Refinement): The set S(v) of source views added to each group is described only as “favorable for geometry reconstruction”; later the implementation simply states “We select 4 source views.” The precise selection criterion (angular threshold, overlap score, COLMAP co-visibility, etc.) is load-bearing for the claim that orientation partitioning improves multi-view consistency, yet is not specified. Provide the exact algorithm so that the partitioning can be reproduced.","section":"§3.1"}],"minor_comments":[{"comment":"Fig. 4 caption and surrounding text: the visual comparison of partitions is helpful, but the dashed lines of CityGaussian-V2 are hard to distinguish in grayscale; add a legend or color-code the blocks more clearly.","section":"Fig. 4"},{"comment":"§3.2 Hole Detection: the entropy threshold τ=0.7 is justified by the distribution in Fig. 8, yet the figure itself is never referenced in the main text of the ablation; a short pointer would help readers.","section":"§3.2 / Fig. 8"},{"comment":"Table 3: training times are given only as totals; reporting per-group wall-clock and peak GPU memory would make the efficiency claim more informative.","section":"Table 3"},{"comment":"UrbanScene3D results (Fig. 7) are purely qualitative; even a simple Chamfer or visual completeness score against the provided meshes would strengthen the multi-dataset claim.","section":"§4.4 / Fig. 7"},{"comment":"Minor typos: “donates” → “denotes” (§3.1); “theshold” → “threshold” (Table 5 caption); inconsistent hyphenation of “point cloud” / “point-cloud”.","section":"throughout"}],"recommendation":"minor_revision","confidential_remarks":"The discrepancy with CityGS-X numbers is the most worrying practical issue; if the authors’ re-implementation is simply weaker, the ranking may reverse once a correct baseline is used. I would ask the authors to share evaluation scripts privately with the AE before acceptance. Otherwise the technical contribution is solid enough for a solid conference/journal track."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is the partitioning. Instead of the usual spatial blocks that mix opposing cameras, they cluster by viewing direction (cosine DBSCAN, 45°), then split oversized groups by position (K-means, M_max=200) and expand each group with a few source views for multi-view losses. That is a clean departure from VastGaussian / CityGaussian / CityGS-X style cuts, and it matches the geometry intuition that high-overlap, similar-angle views give better depth. Selective hole filling via COLMAP-keypoint entropy + a pretrained matcher is secondary but practical; it only densifies the sparse regions that actually hurt 3DGS initialization.\n\nWhat they show is solid for the subfield. Mean F1 on GauU-Scene is 0.560 vs 0.506 / 0.527 for the two strongest baselines; MatrixCity jumps to 0.803 vs 0.556 / 0.581. Visuals on UrbanScene3D look cleaner too. Ablations (Tables 4–5) are honest: 45° is best, but 15–60° still beats or matches CityGS-X, and even the no-completion row stays competitive. Training times are reasonable (2 h / 6.5 h on 8×3090). They reuse PGSR losses and report the usual PSNR/SSIM/LPIPS in the supplement; citations cover the right large-scale 3DGS and surface papers.\n\nSoft spots are the ordinary CV ones, not load-bearing. Thresholds are hand-set (though ablated), GT LiDAR is cropped to central regions following CityGaussian-V2, CityGS-X numbers differ from the original paper, and there are no error bars or code yet. None of that invents the gains or creates circularity—metrics are against external point clouds. The central claim survives the stress test.\n\nThis is for people already doing city-scale 3DGS surfaces or digital-twin pipelines. It will not rewrite theory, but it is a clear engineering step that a serious referee should see. I would send it out; if code appears it becomes immediately usable.","headline":"Orientation-first view clustering is a real, simple fix for large-scale 3DGS surfaces; gains hold even when thresholds move and completion is ablated.","tokens_in":15816,"tokens_out":543,"would_cite":true,"duration_ms":5023,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Grouping city-scale camera views by orientation, not space, plus filling sparse point-cloud holes yields more accurate large-scene surfaces than prior 3DGS methods.","keywords":["3D surface reconstruction","large-scale scenes","3D Gaussian Splatting","viewpoint orientation partitioning","point-cloud completion","multi-view consistency","city-scale geometry"],"falsifier":"Re-run the identical training schedule on GauU-Scene and MatrixCity while sweeping the orientation threshold away from 45° (or the entropy threshold away from 0.7); if mean F1 falls below the CityGS-X baseline for most settings, the claim that these fixed thresholds are reliable collapses.","tokens_in":15930,"feed_emoji":"🏙️","tokens_out":577,"duration_ms":5587,"temperature":0.7,"pith_summary":"Large outdoor and urban scenes still produce coarse or incomplete surfaces under 3D Gaussian Splatting because spatial blocks mix opposing camera directions and because Structure-from-Motion leaves holes in poorly textured or sparsely observed regions. This paper claims that clustering cameras first by viewing direction (cosine distance) and then by position produces viewpoint-consistent groups that share high overlap, so multi-view photometric and geometric constraints recover accurate depth. A second, selective completion step detects images whose projected keypoints have low spatial entropy and densifies only those holes with a pretrained matcher. The resulting pipeline runs in parallel across ordinary GPUs and reports higher surface F1 scores than recent city-scale Gaussian methods on three public benchmarks. A reader who needs reliable city meshes for mapping or simulation cares because the gains come from a simple reordering of existing views rather than heavier models or denser captures.","feed_headline":"Orientation-based view groups beat spatial blocks for city meshes","feed_subtitle":"Clustering cameras by look direction plus selective hole filling raises surface F1 on three urban benchmarks.","key_machinery":"Viewpoint Orientation Partitioning (VOP): two-stage clustering that first groups cameras by cosine distance of viewing directions (threshold ≈45°) then splits oversized groups by 3D position; source-view refinement and multi-view consistency losses are applied only inside each orientation-consistent group.","core_discovery":"Viewpoint-orientation partitioning—DBSCAN on camera directions followed by position K-means that caps each group at 200 cameras—plus entropy-guided completion of COLMAP holes produces higher-fidelity surfaces than spatial-block 3DGS baselines, because views with similar orientations jointly constrain depth more reliably than views that merely lie near one another in space.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Camera direction clusters beat spatial blocks for city surfaces","Orientation partitioning plus hole fill yields sharper urban meshes","Similar-view groups constrain depth better than nearby cameras","DBSCAN on orientations plus completion tops spatial 3DGS baselines","Viewpoint clusters raise city-scale surface F1 over block methods"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single fixed orientation threshold of roughly 45 degrees and a single entropy threshold of 0.7 on COLMAP keypoints correctly identify both the best view groups and the true geometric holes across every tested city-scale scene.","fun_headline_variants_meta":{"raw":{"variants":["Camera direction clusters beat spatial blocks for city surfaces","Orientation partitioning plus hole fill yields sharper urban meshes","Similar-view groups constrain depth better than nearby cameras","DBSCAN on orientations plus completion tops spatial 3DGS baselines","Viewpoint clusters raise city-scale surface F1 over block methods"]},"model":"grok-4.5","effort":"low","cost_usd":0.0053,"raw_usage":{"total_tokens":1446,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":53000000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":617,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":81,"duration_ms":4823,"temperature":1.0,"reasoning_tokens":617,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:03:22.447460+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the identical training schedule on GauU-Scene and MatrixCity while sweeping the orientation threshold away from 45° (or the entropy threshold away from 0.7); if mean F1 falls below the CityGS-X baseline for most settings, the claim that these fixed thresholds are reliable collapses.","supporting_citations":[],"review_version":1}