{"id":"224119df-8852-46f4-b038-0b9d22d08f70","arxiv_id":"2412.01402","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ULSR-GS partitions large aerial scenes by sparse SfM points and trains each sub-region with multi-view geometric consistency, improving surface mesh accuracy.","lead":"A method called ULSR-GS uses the 3D point cloud of a scene, not just camera positions, to split large aerial photos into smaller regions and then trains Gaussian splatting models that enforce multi-view depth and normal consistency. It aims to make city-scale 3D surface reconstruction from drone photos faster and more detailed than existing Gaussian splatting or MVS pipelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"F1 results are computed after cropping to each mesh's overlap with LiDAR, so ULSR-GS's known background omission (Table 7) is excised before scoring; rankings may reflect evaluation extent, not geometric accuracy.","rationale":"The reader's weakest_assumption correctly flags the evaluation protocol in Sec. 5.3.2; I sharpen it: the issue is not primarily ICP residual tolerance but the overlap-based cropping, which changes the evaluation domain per method and systematically removes the regions ULSR-GS intentionally discards. Table 7 is an in-paper admission that the partitioning deletes background regions, and Sec. 4.1.5 describes the density filter that drops sparse sub-regions. Since each method's mesh has a different extent, cropping to the per-method/GT intersection means Table 2's F1 values are not computed on a common scene volume; this is an internally inconsistent comparison even before considering alignment. This directly threatens the strongest_claim, which is about accuracy on complex urban scenes, not about accuracy on the intersection sub-volume. I did not select the potential mathematical issue in Eq. (16) as the primary concern because even if that loss formulation is loose, the method's empirical ranking could still hold; the evaluation-extent issue, by contrast, can change the ranking itself. The paper does have real strengths: external LiDAR GT, a clear ablation showing the losses and densification matter, and a well-described pipeline. Those strengths are why the appropriate response is to demand a corrected evaluation rather than to reject outright. A conditional verdict with the full-extent recomputation as the condition is the right outcome.","tokens_in":23665,"tokens_out":6978,"duration_ms":64979,"concrete_test":"Recompute Tables 2, 3, and 4 using the full LiDAR ground-truth bounding box as the fixed evaluation volume for all methods, without cropping each mesh to its own overlap; as a robustness check, also compute with the union bbox and with per-subregion masks. Report ICP residual RMSE before and after alignment for every scene, plus per-subregion seam error statistics along shared tile boundaries. If ULSR-GS retains its F1 lead over PGSR on all six GauU scenes and over CityGaussianV2 on Matrix City under full-extent evaluation, the current numbers are validated; if the lead shrinks or flips, the central claim overstates the improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on precision/recall/F1 in Tables 2-4, computed per Sec. 5.3.2. Step 1 crops both the reconstructed mesh and the LiDAR ground truth to their overlapping axis-aligned bounding box before ICP and before any metric is computed. This is not a neutral common evaluation domain: each method has a different mesh extent, so each method's F1 is evaluated on a different intersection volume. The effect is not hypothetical. ULSR-GS's partitioning (Sec. 4.1.5) explicitly discards sub-regions with sparse points, and Table 7 concedes that the strategy 'deletes the background area' and lowers rendering quality on background regions. Under the overlap-cropping protocol, exactly those omitted background GT points are removed from the recall denominator, and any false-positive mesh geometry outside the GT bbox is removed from the precision numerator. A method that reconstructs only a small, well-aligned core can obtain a high F1 while missing large portions of the scene; because the cropped domain differs per method, the across-method comparisons in Tables 2-4 are not made on the same region. The ICP step adds a second, unquantified degree of freedom: no residuals are reported, so it is unclear whether tau=0.025 (GauU) or tau=1.0/0.5 m (Scene1/2) is large relative to residual misalignment. Per-subregion seam errors from stitching are also never reported. Until F1 is recomputed on a fixed evaluation extent (e.g., full LiDAR bbox or union bbox) and with ICP residuals reported, the headline 'significantly improving surface extraction accuracy' is not established by the numbers as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ULSR-GS, a large-scale surface reconstruction framework built on 2D Gaussian Splatting. Its three main components are a point-to-photo scene partitioning strategy driven by sparse SfM points, a per-point optimal view selection that follows the MVS-Net scoring formulation, and an adaptive multi-view geometric-consistency densification with depth and normal consistency losses. The method is evaluated on the GauU-Scene benchmark, Matrix City, and two custom-collected aerial datasets, comparing against single-GPU GS baselines (SuGaR, 2DGS, GOF, PGSR), a multi-GPU baseline (CityGaussianV2), and MVS pipelines (COLMAP, Reality Capture), using precision, recall, and F1 against LiDAR ground truth. The central claim is that ULSR-GS outperforms existing GS-based surface extraction on large-scale aerial photogrammetry while being more time-efficient than MVS.","tokens_in":23944,"tokens_out":5546,"duration_ms":45429,"significance":"If the reported results are valid, ULSR-GS is a practically relevant contribution to city-scale photogrammetry: it extracts meshes per sub-region without merging the entire scene first, and its ablation table is consistent with the individual contributions of the proposed modules (removing normal consistency lowers F1 from 0.685 to 0.651, removing densification lowers it to 0.682). Strengths include evaluation against external LiDAR on six GauU scenes, a multi-GPU pipeline with per-tile training, and a clear qualitative comparison. The main quantitative claim, however, rests on an evaluation protocol that crops each mesh to its own overlap with LiDAR; because ULSR-GS deliberately discards sparse background sub-regions, this protocol can reward selective reconstruction and confound cross-method comparisons. The headline results are therefore not yet fully supported.","major_comments":[{"comment":"The mesh evaluation protocol crops both the reconstructed mesh and the LiDAR ground truth to their overlapping axis-aligned bounding box before any metric is computed. Since each method has a different mesh extent, precision, recall, and F1 are computed on different domains for different methods. Sections 4.1.5 and Table 7 explicitly state that ULSR-GS discards sparse background sub-regions and that the partition strategy 'deletes the background area.' Under the overlap-cropping protocol, the omitted background LiDAR points are removed from the recall denominator for ULSR-GS, and any out-of-bbox mesh geometry is removed from the precision numerator. This makes the F1 rankings in Tables 2-4 reflect evaluation extent as well as geometric accuracy. The authors should recompute all metrics on a fixed evaluation domain (e.g., the full LiDAR bounding box or the union of all compared mesh bounding boxes), report the fraction of LiDAR points inside each mesh's overlap, and report ICP residuals and the sensitivity of F1 to the alignment threshold. Without this, the central comparison claim is not established.","section":"Sec. 5.3.2, Tables 2-4"},{"comment":"The comparison with single-GPU baselines is not hardware-matched. Ours trains each sub-region on one of four RTX 4090 GPUs, while SuGaR, 2DGS, GOF, and PGSR run on a single GPU at 60k iterations. The wall-clock time column in Table 2 for Ours therefore does not reflect the roughly four times higher GPU-hour consumption. The claimed efficiency advantage over GOF and PGSR should be expressed in GPU-hours, with the iteration budget and training resolution per sub-region specified. If the intended claim is quality at comparable compute, the baselines should be run under a comparable multi-GPU setting, or the text should restrict the claim to end-to-end wall-clock time on the authors' specific hardware configuration.","section":"Sec. 5.5.1, Table 2"},{"comment":"The adaptive densification window size is defined as (h_win, w_win) = k / (g_bar + epsilon) * (h, w) / 2, but the values of k and epsilon are never reported in Section 5.1 or in the ablations. Since the window mask is essential to avoid GPU out-of-memory (as shown by the ablation 'w densify & w/o window mask' in Table 5) and directly controls the number of added Gaussians, the choice of k and epsilon materially affects the densification outcome. Please report the values used and, if possible, provide a sensitivity analysis.","section":"Sec. 4.2.2, Eq. (18)"},{"comment":"The threshold 'tau = 0.025 (relative)' is not defined. The Tanks and Temples pipeline normally uses absolute thresholds in scene units; the paper should state how 'relative' is computed per scene (e.g., as a fraction of the bounding-box diagonal or mean scene scale). Without this, the F1 numbers in Table 2 are not reproducible, and it is unclear whether the same physical tolerance is used across the six GauU-Scene sites.","section":"Sec. 5.3.2, Table 2"}],"minor_comments":[{"comment":"There is a typo 'we we calculate' in the description of the average Euclidean distance, and 'agrmin' in Eq. (12) should be 'argmin'.","section":"Sec. 4.1.4"},{"comment":"Several figure captions contain typos, including 'cumtom' for 'custom' in Figures 8-9, 'ars' for 'are' in Figures 14-15, and 'Rensults' for 'Results' in Figure 5.","section":"Figure captions"},{"comment":"The same paper appears twice in the reference list: VastGaussian is listed as [29] and again as [30] with identical metadata; the in-text citations should be unified.","section":"References"},{"comment":"The symbols L_geo and L_r in the final loss are not explicitly defined; the text later describes them as the 2DGS and 3DGS losses, but the equation should carry a clear definition.","section":"Sec. 4.3, Eq. (21)"},{"comment":"The sentence 'We uniformly sample number of points' is missing an article; it should read 'We uniformly sample the number of points'.","section":"Sec. 5.3.2"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-extent issue is the main obstacle to publication. Because ULSR-GS explicitly discards sparse background regions (Table 7), the overlap-cropping protocol in Sec. 5.3.2 may systematically inflate its F1 scores relative to methods that reconstruct the full scene. I recommend that the editor ask for a fixed-extent re-evaluation and for ICP residual reporting before considering acceptance. The duplicate VastGaussian citation should be fixed. The paper should also clarify how closely the densification and loss components follow the authors' prior MVG-Splatting [27], since several steps are described as 'following' that work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nULSR-GS is a genuine systems contribution: point-to-photo partitioning based on per-point optimal view groups, adaptive window-masked multi-view densification, and two multi-view consistency losses on 2DGS sub-regions. The writing is clear, the math in Sec. 4 is consistent, and the ablation (Table 5) supports the claimed function of each module: removing normal consistency drops F1 from 0.685 to 0.651; removing densification drops it to 0.682. The geometry evaluation is against external LiDAR, so circularity is low. That is real credit.\n\nThe soft spots are real and mostly about evaluation rather than method logic. The largest is the metric protocol in Sec. 5.3.2: each mesh is cropped to its own overlapping AABB with the LiDAR before ICP and F1. Since ULSR-GS's partitioning explicitly deletes sparse background sub-regions (Table 7 concedes this), the method is scored on a smaller, denser core while baselines are scored on their own extents. The stress-test note is right: the across-method F1 rankings in Tables 2-4 are not computed on a common evaluation domain. The paper should report F1 on a fixed bbox (e.g., full LiDAR bbox or union bbox) and give ICP residuals and per-block seam errors. Until then I would treat the 'significantly improving' claim as conditional.\n\nOther issues are secondary. The custom Scene1/Scene2 datasets have no public data or code, which is where the strongest runtime claims live. Hyperparameters like theta_0, k, epsilon in Eq. 18, alpha/beta are given but without sensitivity. And the single-GPU baseline comparison is not apples-to-apples: 4 RTX 4090s vs 1 GPU, with iteration budgets adjusted in a way that can favor the partitioned method. These are fixable with more transparency.\n\nWho is the paper for? Photogrammetry and large-scale GS practitioners who want a scalable mesh-extraction recipe. The reader gets a coherent, well-tested system with honest ablations, plus a clear warning about evaluation pitfalls. I would send this to peer review: the engineering is substantive enough to warrant referee time, and the evaluation flaw is testable and fixable rather than fatal. After a revision that recomputes the numbers on a fixed evaluation extent and releases code/data, the central claims would likely hold.","headline":"Solid large-scale GS surface-reconstruction system whose F1 claims are undercut by per-method evaluation cropping; worth reviewing after the metric protocol is fixed.","tokens_in":24629,"tokens_out":2994,"would_cite":false,"duration_ms":26007,"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's central claim is that Gaussian-splatting surface extraction scales to ultra-large aerial scenes when sub-regions are partitioned by SfM points and trained with multi-view depth and normal consistency.","keywords":["Gaussian Splatting","Surface Reconstruction","Large-Scale Scenes","Aerial Photogrammetry","Urban Scene Reconstruction","Multi-View Geometric Consistency","Point Cloud Partitioning"],"falsifier":"Re-run the same scene evaluations with the LiDAR ground truth fixed by surveyed ground-control markers instead of best-fit alignment, or perturb the alignment by a few centimeters and recompute F1; if the reported margins over PGSR and CityGaussianV2 shrink below the perturbation or reverse, the accuracy claim would be measuring alignment quality rather than reconstruction quality.","tokens_in":23371,"feed_emoji":"🏙️","tokens_out":9913,"duration_ms":80102,"temperature":0.7,"pith_summary":"The paper argues that Gaussian Splatting, which already gives fast rendering and small-scene surface extraction, can be pushed to ultra-large-scale aerial photogrammetry if the scene is partitioned the right way. ULSR-GS divides the scene using the sparse SfM point cloud instead of camera positions, then assigns each 3D point the four-image group that sees it closest to the image centers, so every sub-region gets the most informative training views. During training it densifies with multi-view depth aggregation and adds depth- and normal-consistency losses, which preserve fine surface detail that single-view densification washes out. The method extracts each sub-region's mesh independently and stitches them without global fusion, avoiding the memory blow-up of merging the whole scene. On the GauU-Scene, Matrix City, and two custom aerial benchmarks, the paper reports higher F1 than existing GS-based extractors, and comparable F1 to MVS pipelines at a fraction of the runtime.","feed_headline":"City-scale 3D meshes from Gaussian splatting at a fraction of MVS time","feed_subtitle":"Point-based partitioning plus multi-view consistency lifts F1 on aerial benchmarks and cuts runtime from days to hours.","key_machinery":"The machinery has three coupled parts. First, the point-to-photo partition: after density filtering of the SfM cloud, the scene is split into rectangular sub-regions, and for every point the method picks the group of one reference plus three source views whose projections lie closest to the image centers, discarding images that no point selects. Second, adaptive multi-view densification: rendered depths from the three source views are fused with geometric-consistency weights and reprojected as new 2D Gaussian primitives only inside an adaptive window mask whose size shrinks where the depth gradient is high. Third, two consistency losses, a depth reprojection error and a normal-vector angular error between reference and source views, push the trained surface to agree across views. Per-region TSDF fusion with boundary cropping then extracts meshes that stitch together without a global fusion step.","core_discovery":"On the paper's own terms, the central discovery is that the bottleneck in large-scale Gaussian surface reconstruction is not the splatting itself but the way the scene is split and supervised. ULSR-GS shows that a point-to-photo partition, in which each SfM point selects its most central four-view group, produces sub-regions dense and consistent enough for 2D Gaussian Splatting to recover fine urban geometry locally, and that the local meshes tile into a coherent whole without merging. Multi-view depth and normal consistency constraints are what carry the improvement: in the ablation, removing both drops F1 from 0.685 to 0.660, while removing only the normal term gives 0.651 and removing only the depth term gives 0.664. The paper's headline numbers include F1 0.718 on SZTU versus 0.672 for PGSR, F1 0.892 on Matrix City versus 0.886 for CityGaussianV2, and competitive F1 with COLMAP on Scene 1 while taking roughly a third of the runtime.","pith_inferences":["One consequence the paper leaves implicit is that its per-point optimal view selection is a general principle: any reconstruction pipeline that starts from SfM tracks could adopt the same closest-to-principal-point four-view group rule to improve its training-view selection.","The ablation pattern hints that densification is not the primary source of geometric accuracy; if so, pairing the consistency losses with cheaper or learned densification might recover most of the benefit at lower memory cost, which the paper does not test.","The method's stated limit near 1.6K rendering resolution and its weakness on vegetation suggest the next testable extension is a higher-resolution rasterizer plus a semantic or depth-prior branch for foliage, which the paper lists as future work.","The evaluation's reliance on best-fit alignment after bounding-box cropping means the reported F1 gaps should be re-checked with survey-registered ground truth; that check is not in the paper but is directly enabled by the evaluation protocol it describes."],"forward_implications":["GS-based mesh extraction can scale to scenes with thousands of oblique aerial images by training sub-regions independently on multiple GPUs, without building a whole-scene dense point cloud or mesh.","View selection should be driven by the geometry being reconstructed from SfM points and their optimal image groups, not by camera positions; this matters most for irregular and close-range flight paths.","Multi-view depth and normal consistency, not densification alone, are the main levers for surface fidelity in large urban scenes; the normal-consistency term is the more influential of the two in the ablation.","Compared to MVS, a GS pipeline can deliver comparable precision and recall on city-scale meshes in about a third of the time, for instance Scene 1 at F1 0.727 in 19.8 hours versus COLMAP at 0.756 in 69.2 hours.","Sub-region TSDF extraction followed by direct stitching is sufficient to avoid seams at sub-region boundaries, provided each region starts with doubled boundaries and redundant points are pruned by opacity culling."],"supporting_citations":[{"why":"Supplies the 2D Gaussian Splatting base representation, the depth aggregation formulas, and the geometric and distortion regularization losses ULSR-GS trains on.","marker":"[16]"},{"why":"Supplies the 3D Gaussian Splatting rendering and optimization backbone, including the L1 and D-SSIM image losses reused in the final loss.","marker":"[19]"},{"why":"Provides the COLMAP SfM pipeline whose sparse point cloud, camera poses, and reprojection errors drive the point-to-photo partitioning and initial view selection.","marker":"[44]"},{"why":"Supplies the multi-view pair selection scoring with baseline angles and the geometric-consistency-weighted depth aggregation that the densification and consistency losses build on.","marker":"[61]"},{"why":"Provides the depth-projection densification mechanism, primitive rescaling and rotational alignment, and the adaptive quantization precedent that the window-mask densification adapts.","marker":"[27]"},{"why":"Supplies the Tanks and Temples evaluation pipeline used to compute precision, recall, and F1 against LiDAR ground truth.","marker":"[21]"},{"why":"Provides the GauU-Scene benchmark with LiDAR ground truth for the single-GPU comparisons and the main ablation experiments.","marker":"[54]"},{"why":"Provides the Matrix City small-city dataset used for the multi-GPU comparison and the 1.0-meter and 0.5-meter threshold evaluations.","marker":"[25]"},{"why":"Supplies the VastGaussian image-position-based partitioning strategy and evaluation protocol that ULSR-GS compares against and improves on.","marker":"[29]"},{"why":"Provides CityGaussianV2, the multi-GPU Gaussian baseline whose partition method and mesh results ULSR-GS claims to outperform on large scenes.","marker":"[34]"}],"fun_headline_variants":["Gaussian splatting goes city-scale with smart point-to-photo splits","Urban meshes from Gaussian splats: partitioning is the key","Multi-view consistency refines city-scale Gaussian surfaces","City-scale Gaussian meshes: better F1, a third of the time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy numbers assume the laser-scanned ground truth and the reconstructed mesh can be aligned well enough by a best-fit geometric alignment after cropping to their overlap, and that the resulting precision, recall, and F1 therefore measure surface fidelity rather than registration error.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian splatting goes city-scale with smart point-to-photo splits","Urban meshes from Gaussian splats: partitioning is the key","Multi-view consistency refines city-scale Gaussian surfaces","City-scale Gaussian meshes: better F1, a third of the time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000728,"raw_usage":{"total_tokens":3256,"prompt_tokens":933,"completion_tokens":2323,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":2249}},"tokens_in":549,"tokens_out":2323,"duration_ms":15881,"temperature":1.0,"reasoning_tokens":2249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:23:17.515339+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same scene evaluations with the LiDAR ground truth fixed by surveyed ground-control markers instead of best-fit alignment, or perturb the alignment by a few centimeters and recompute F1; if the reported margins over PGSR and CityGaussianV2 shrink below the perturbation or reverse, the accuracy claim would be measuring alignment quality rather than reconstruction quality.","supporting_citations":[{"cited_title":"3d gaussiansplattingforreal-timeradiancefieldrendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting rendering and optimization backbone, including the L1 and D-SSIM image losses reused in the final loss."},{"cited_title":"Pixel- wise view selection for unstructured multi-view stereo, in: European Conference on Computer Vision (ECCV)","cited_arxiv_id":null,"evidence_quote":"Provides the COLMAP SfM pipeline whose sparse point cloud, camera poses, and reprojection errors drive the point-to-photo partitioning and initial view selection."},{"cited_title":"Mvsnet: Depth inference for unstructured multi-view stereo, in: Proceedings of the European conference on computer vision (ECCV), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-view pair selection scoring with baseline angles and the geometric-consistency-weighted depth aggregation that the densification and consistency losses build on."},{"cited_title":"Tanks and temples: Benchmarking large-scale scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the Tanks and Temples evaluation pipeline used to compute precision, recall, and F1 against LiDAR ground truth."},{"cited_title":"Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond","cited_arxiv_id":null,"evidence_quote":"Provides the Matrix City small-city dataset used for the multi-GPU comparison and the 1.0-meter and 0.5-meter threshold evaluations."},{"cited_title":"Vastgaussian: Vast 3d gaussians for large scene reconstruction, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the VastGaussian image-position-based partitioning strategy and evaluation protocol that ULSR-GS compares against and improves on."},{"cited_title":"Citygaussianv2: Efficient and geometrically accurate reconstruction for large-scale scenes, in: ICLR","cited_arxiv_id":null,"evidence_quote":"Provides CityGaussianV2, the multi-GPU Gaussian baseline whose partition method and mesh results ULSR-GS claims to outperform on large scenes."}],"review_version":1}