{"id":"c80da686-5085-4c78-b8a4-71716e6a5fcb","arxiv_id":"2501.01695","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CrossView-GS reconstructs large-scale 3D scenes from cross-view images by training branch models as priors, applying gradient-aware regularization that preserves salient gradients, and supplementing the cross-view model with unique Gaussian primitives from each branch.","lead":"This paper proposes a multi-branch 3D Gaussian Splatting method to reconstruct large-scale scenes from images with very different viewpoints, such as aerial and ground captures. It addresses a common failure of existing techniques that occurs when view differences are large, making the method useful for drone and mobile mapping applications.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own numbers undermine an unqualified 'superior performance' claim: CrossView-GS has higher LPIPS than Horizon-GS on 3 of 4 aerial-ground scenes (Block small, NYC, SF) in Table I, so the central claim needs qualification or statistical backing.","rationale":"The reader's weakest_assumption was the voxel-grid alignment in the unique Gaussian supplementation (Eq. 6). That is an implementation detail that, if wrong, would affect the described mechanism but would not directly falsify the headline empirical claim; the reported gains could still arise from the other components. The LPIPS inconsistency, by contrast, is a direct contradiction in the paper's own quantitative results: the abstract makes a universal superiority claim, yet the strongest baseline wins the perceptual metric on most scenes. This is the more load-bearing concern because it targets the central claim itself and can be settled by standard statistical re-evaluation. I therefore partially disagree with the reader's choice of weakest assumption, while agreeing with the overall conditional verdict.","tokens_in":12029,"tokens_out":13175,"duration_ms":127369,"concrete_test":"Run 5 independent training runs of CrossView-GS and Horizon-GS on NYC and SF with the same training/test split and the authors' hyperparameters, then report mean and 95% confidence intervals for PSNR, SSIM, and LPIPS. If LPIPS remains significantly worse for CrossView-GS, the abstract's 'superior performance' claim should be revised; if the intervals overlap, the current claim is not statistically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CrossView-GS 'achieves superior performance in novel view synthesis compared to state-of-the-art methods' (Abstract). For this to hold, the method should be better than the strongest baseline across the reported metrics. Table I (Section V-A) shows the opposite on LPIPS: versus Horizon-GS, CrossView-GS scores 0.213 vs 0.197 on Block small, 0.199 vs 0.189 on NYC, and 0.269 vs 0.240 on SF, and only wins on Zeche (0.201 vs 0.215). LPIPS is a standard perceptual metric for view synthesis, so an unqualified superiority claim is contradicted by the authors' own results. Further, no error bars or multiple seeds are provided; on SF the PSNR advantage is just +0.04 dB (27.702 vs 27.663), within typical run-to-run noise. Thus the main claim rests on favorable interpretation of PSNR/SSIM while ignoring LPIPS regressions. The paper should either qualify the claim (e.g., 'better PSNR/SSIM, comparable or worse LPIPS') or provide statistical evidence that the aggregate improvement is significant despite the LPIPS losses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CrossView-GS, a 3D Gaussian Splatting method for reconstructing large-scale scenes from cross-view image sets, such as aerial and ground views. The method has three components: (1) multi-branch initialization, where sub-models are trained on separate view sets and the distant-view sub-model's Gaussians are downsampled to produce the initial point cloud for a cross-view model; (2) gradient-aware regularization, which modifies the densification criterion to use the maximum per-view gradient (Eq. 4) and adds a pseudo-label loss that penalizes predictions worse than the branch-model renderings (Eq. 5); and (3) unique Gaussian supplementation, which inserts Gaussians from branch models whose anchors lie outside the cross-view model's voxel grid (Eq. 6) and fine-tunes the combined model. Experiments on MatrixCity, UC-GS, ISPRS, and two newly collected internet-video datasets report PSNR, SSIM, and LPIPS against 3DGS, Mip-Splatting, Scaffold-GS, Hier-GS, and Horizon-GS.","tokens_in":12361,"tokens_out":6763,"duration_ms":59729,"significance":"The motivation is sound: cross-view data violate the smooth-view assumption of adaptive densification control in 3DGS, and the paper gives a clear, arithmetically correct derivation of why per-view salient gradients get averaged away (Eqs. 2-4). The proposed max-gradient rule is a simple, plausible fix. The systematic ablation in Table V demonstrates that each component contributes to PSNR/SSIM/LPIPS gains over the Scaffold-GS baseline, and the inclusion of real captured (internet) datasets and a limitations section (dynamic objects) are positives. However, the unqualified claim of 'superior performance' in the abstract and conclusion is undercut by the paper's own LPIPS numbers in Table I, where CrossView-GS is worse than Horizon-GS on three of four aerial-ground scenes. No error bars or multi-seed statistics are reported, and the evaluation protocol for the UC-GS-derived scenes is incompletely specified. These issues must be resolved before the central claim can be accepted.","major_comments":[{"comment":"The abstract and conclusion claim 'superior performance' in novel view synthesis compared to state-of-the-art methods. However, Table I shows that CrossView-GS has higher (worse) LPIPS than Horizon-GS on Block small (0.213 vs. 0.197), NYC (0.199 vs. 0.189), and SF (0.269 vs. 0.240), and it ties with Horizon-GS on SSIM on Block small (0.846) while being worse on SF SSIM (0.784 vs. 0.791). Since LPIPS is a standard perceptual metric reported by the paper itself, the unqualified superiority claim is not supported. Please either qualify the claim (e.g., 'better PSNR/SSIM, comparable or worse LPIPS'), or provide multi-seed runs with error bars and a statistical test demonstrating that the aggregate improvement is significant despite these regressions.","section":"Abstract; Table I"},{"comment":"The quantitative comparison for the UC-GS-derived NYC and SF scenes in Table I uses a newly introduced split (one test image out of every eight based on image suffix number). The paper does not state whether the baseline results in Table I were obtained by retraining each baseline on this new split or were copied from the original papers. If baselines were not retrained on this split, the comparison is unfair and the reported margins may not be meaningful. Please specify the exact split, describe how each baseline was adapted to it, and, ideally, release the train/test index files for reproducibility.","section":"Section V-A, Table I"},{"comment":"The unique Gaussian supplementation step defines non-overlap as νi ∉ Vf , where Vf is the voxel grid of the cross-view model and νi is the anchor voxel of a Gaussian from a branch model. This test is only valid if the branch-model grids and the cross-view model grid share the same coordinate frame, voxel origin, and lattice. The paper only states that the models are 'reconstructed using Scaffold-GS with the same voxel size.' Since Scaffold-GS anchors are created during training and depend on the initial point cloud and densification history, and the branch models and cross-view model are initialized from different point clouds, the grids are not guaranteed to be aligned. If they are misaligned, the supplementation will insert redundant or displaced Gaussians and the method's reported gains may not transfer to other scenes. Please specify the alignment procedure or provide evidence (e.g., a quantitative analysis of grid overlap) that the grids are co-registered.","section":"Section IV-C, Eq. (6)"},{"comment":"The regularization loss Lreg uses renderings from the multi-branch sub-models as pseudo-labels for the training views. Because these sub-models are trained on the same training images that are later used to train the cross-view model, this is a self-distillation mechanism. If a sub-model's rendering is erroneous on a given view (e.g., due to incomplete geometry or view-specific artifacts), the loss penalizes the cross-view model for deviating from that error, potentially reinforcing it. The paper does not discuss this failure mode or provide evidence that the pseudo-labels are reliable, such as an ablation that replaces the pseudo-label with the ground truth or a comparison of pseudo-label quality on held-out views. Please add such an analysis or justify why the risk is negligible.","section":"Section IV-B, Eq. (5)"}],"minor_comments":[{"comment":"The statement 'All GS-based methods are initialized with the same point cloud' is ambiguous: does this mean the baselines also use the point cloud derived from the distant-view sub-model, or that all methods use the same COLMAP point cloud? This is important because the proposed initialization alone yields a large gain (Table V, 'w/ Init.'). Please clarify.","section":"Section V, Baselines"},{"comment":"The table caption notes that the SOTA methods are trained for 900k iterations while the proposed method uses 30k+20k. This is a strength, but the iteration counts should be stated directly in the table or caption so readers can verify the comparison without reading the text.","section":"Table IV"},{"comment":"The caption says 'variation of maximum gradient in different views' but the axes are not labeled. Please label the axes and clarify what is plotted (e.g., maximum gradient over iterations).","section":"Figure 3"},{"comment":"References [5] and [6] are incompletely specified: [5] is a bare arXiv identifier with no year/venue, and [6] is an arXiv preprint; please complete the bibliographic information for these entries.","section":"References"},{"comment":"The limitations section mentions dynamic objects but does not discuss the effect of the newly introduced split or the lack of code; consider adding a reproducibility statement.","section":"Section V-D"},{"comment":"The dataset collection for Temple and Monument relies on MASt3R-SfM for pose estimation; please report the pose accuracy or registration errors for these datasets, since the cross-view alignment is critical to the method.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"I would like to flag that the paper does not release code or the newly collected datasets, which limits reproducibility and verification of the claimed gains. Also, the LPIPS regressions in Table I are not addressed in the paper's discussion; the authors should be asked to confront this issue head-on. The paper fits the journal's scope but the evaluation protocol needs to be tightened before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does one thing well and overclaims one thing. The gradient-averaging diagnosis is clean, and the max-gradient densification rule is a sensible fix. But the abstract's \"superior performance\" does not survive contact with Table I: on 3 of 4 aerial-ground scenes, LPIPS is worse than Horizon-GS, and on SF the PSNR gain is only 0.04 dB.\n\nThe genuinely new piece is the derivation in Eqs. (2)-(4): averaging gradients across very different views smooths out the salient gradients needed for densification, and taking the max per-view gradient is a natural correction. The multi-branch initialization from distant views plus the supplementation of unique Gaussians is a reasonable architecture, and the ablation shows each component adds something. The experiments span aerial-ground, pure aerial, and pure ground settings, and the UC-GS comparison (Table IV) shows competitive results with far fewer iterations than UC-GS's 900k. That is a meaningful practical contribution, and the paper is clearly written.\n\nThe soft spots are real but fixable. First, the LPIPS regressions are not minor: on Block small, NYC, and SF, Horizon-GS is better on LPIPS by 0.016, 0.010, and 0.029 respectively. The abstract needs qualification, or the paper needs statistical evidence that the PSNR/SSIM gains are consistent despite the perceptual losses. Second, there are no error bars, and the SF PSNR advantage is within run-to-run noise. Third, the voxel-alignment assumption in Sec. IV-C is under-specified: \"same voxel size\" does not guarantee aligned grids, since Scaffold-GS places voxels based on the initial point cloud and training. If the grids are misaligned, the supplementation could add redundant or displaced Gaussians. The pseudo-label regularization is mild self-distillation, but it is only a regularizer and the ablations show gains, so I would not call it circular.\n\nBottom line: this is a serious paper worth engaging with. A good referee should ask for qualification of the central claim, error bars, code release, and alignment details. With those, it would be a solid contribution to cross-view 3DGS. I would send it to peer review.","headline":"A solid cross-view 3DGS method with a clear gradient-smoothing diagnosis, but the 'superior performance' claim is undercut by the paper's own LPIPS numbers.","tokens_in":12818,"tokens_out":1758,"would_cite":true,"duration_ms":16628,"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":"A multi-branch Gaussian splatting pipeline overcomes the gradient smoothing that limits cross-view scene reconstruction, delivering higher-quality novel views than prior state-of-the-art methods on aerial-ground, pure-aerial, and…","keywords":["3D Gaussian Splatting","cross-view reconstruction","novel view synthesis","large-scale scene reconstruction","adaptive densification control","multi-branch fusion","gradient-aware regularization"],"falsifier":"A direct test would be to train a cross-view scene twice: once with the stated supplementation and once after applying a small rigid perturbation (for example a 10 cm translation) to the branch model coordinates before the voxel-overlap test of Eq. (6), while keeping the rest of the pipeline identical. If the perturbed version matches or exceeds the PSNR of the unperturbed version on held-out aerial or ground views, then the precise voxel alignment is not what drives the gains, and the supplementation mechanism would need a different explanation.","tokens_in":11870,"feed_emoji":"🏙️","tokens_out":7211,"duration_ms":60921,"temperature":0.7,"pith_summary":"This paper claims that the standard densification rule in 3D Gaussian Splatting fails when training images come from sharply different viewpoints, such as aerial and ground cameras, because averaging gradients across such views washes out the salient gradients that tell the model where to add Gaussian primitives. To fix this, the authors train separate sub-models on each view set, use the distant-view sub-model to initialize the cross-view model, replace the average-gradient densification test with a per-view maximum-gradient test, and add a regularization term that penalizes regions where the cross-view model renders worse than the branch models. They then fuse Gaussians that each branch created but the cross-view model lacks, judged by a voxel-grid overlap test, and fine-tune briefly. The reported result is consistently higher novel-view quality than prior methods on aerial-ground, pure-aerial, and pure-ground benchmarks, with the largest gains in aerial views and roughly 1.15 dB average PSNR improvement over the Scaffold-GS baseline.","feed_headline":"Cross-view Gaussian splatting gains 1.15 dB by splitting views","feed_subtitle":"A multi-branch pipeline fixes gradient smoothing, lifting aerial and ground rendering quality over prior 3DGS methods.","key_machinery":"The load-bearing mechanism is the max-gradient densification rule of Eq. (4), which replaces the average-gradient test of the original 3DGS densification with a maximum over per-view gradient statistics, so a primitive is densified if any single view set finds it salient. Around this rule the paper builds a three-part pipeline: multi-branch initialization, where sub-models are trained separately on each view set and the distant-view sub-model is downsampled into the starting point cloud; gradient-aware regularization, where branch renderings serve as pseudo-labels in a $\\max(0, d(\\mathrm{pred},\\mathrm{gt})-d(\\mathrm{ref},\\mathrm{gt}))$ loss (Eq. 5) that only penalizes the cross-view model where it is worse than a branch; and unique Gaussian supplementation (Eq. 6), which keeps branch Gaussians whose anchor voxel is absent from the cross-view model and then fine-tunes. The voxel grid shared by the Scaffold-GS models is the object that lets the supplementation test decide which Gaussians are complementary.","core_discovery":"The paper's central claim is that cross-view Gaussian splatting can be made to work by treating each view set as an independent teacher instead of letting all views compete inside one optimization. The authors show that when views with large disparities are pooled, the densification condition $\\frac{\\nabla_1+\\nabla_2}{c_1+c_2}>\\tau$ (Eq. 3) is dominated by the larger, smoother-view gradient statistics, so primitives with salient per-view gradients are never densified. Their remedy, the max-gradient selection $\\max(\\nabla_1/c_1,\\nabla_2/c_2)>\\tau$ (Eq. 4), restores single-view-level densification, and the pseudo-label regularization (Eq. 5) pushes the cross-view model to be at least as good as the best branch. Finally, a voxel-grid membership test (Eq. 6) adds branch-only Gaussians to the cross-view model so fine details from each viewpoint set are retained. The paper argues that this combination yields superior novel-view synthesis, with the largest quality gains on aerial views.","pith_inferences":["The per-view maximum-gradient test in Eq. (4) could be dropped into any 3DGS-style densification routine, not just Scaffold-GS, whenever training views are clustered by appearance or viewpoint; the paper only demonstrates it inside its own pipeline.","The multi-branch pseudo-label loss is effectively a knowledge-distillation signal from view-specific teachers to the fused model, so the same regularization may apply to other cross-domain radiance-field tasks such as day-night or indoor-outdoor reconstruction.","The voxel-alignment assumption suggests a testable boundary: on scenes with noisy or drifting pose estimates, the supplementation step may degrade unless the branch and fused models are re-registered before Eq. (6); enforcing alignment explicitly could make the method more robust."],"forward_implications":["On the tested benchmarks, the method reports higher PSNR, SSIM, and lower LPIPS than 3DGS, Mip-Splatting, Scaffold-GS, Hier-GS, and Horizon-GS, with an average 1.15 dB PSNR gain over Scaffold-GS.","Aerial views benefit most from the gradient-aware regularization and the distant-view initialization, making the method well suited to drone-plus-ground captures.","The same pipeline transfers to pure-aerial and pure-ground pairs with strong view differences, and to three view sets, indicating the approach is not limited to aerial-ground data.","The cross-view model trains in 50k iterations (30k initial plus 20k fine-tuning), far fewer than UC-GS's 900k, while staying competitive or better on the UC-GS benchmark."],"supporting_citations":[{"why":"Backbone model used for all branch sub-models and the cross-view model; its densification rule is the one being modified.","marker":"[1]"},{"why":"Defines the original 3DGS adaptive densification control whose gradient averaging the paper identifies as the source of cross-view smoothing.","marker":"[4]"},{"why":"Supplies the NYC and SF cross-view datasets and the UC-GS baseline used in the main comparison table.","marker":"[6]"},{"why":"Comparison baseline whose similar-to-3DGS results the paper uses to argue the cross-view problem is not a multi-scale sampling issue.","marker":"[13]"},{"why":"State-of-the-art large-scale aerial-to-ground method used as a comparison baseline.","marker":"[23]"},{"why":"Motivates the initialization step where the distant-view sub-model is downsampled into a point cloud for the cross-view model.","marker":"[24]"},{"why":"Provides the max(0, ·) triplet-style formulation adopted for the gradient-aware regularization loss.","marker":"[25]"},{"why":"Supplies the MatrixCity Block small and Block A datasets used for aerial-ground and pure-ground evaluation.","marker":"[26]"},{"why":"Supplies the real Zeche scene used to test the method on actual captured aerial-ground data.","marker":"[27]"}],"fun_headline_variants":["Splitting views boosts 3D Gaussian splatting by 1.15 dB","Cross-view Gaussian splatting with branch experts gains 1.15 dB","Multi-branch Gaussians fix cross-view smoothing, gain 1.15 dB","CrossView-GS: branch-wise Gaussians for large-scale scenes","Gaussian splatting gets smarter across large view gaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the branch sub-models and the cross-view model use identical, aligned voxel grids with the same coordinate frame, so the test $\\nu_i \\notin V_f$ reliably identifies Gaussians that are genuinely complementary; if the grids are even slightly misaligned, the supplementation step can insert redundant or misplaced Gaussians and the reported gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Splitting views boosts 3D Gaussian splatting by 1.15 dB","Cross-view Gaussian splatting with branch experts gains 1.15 dB","Multi-branch Gaussians fix cross-view smoothing, gain 1.15 dB","CrossView-GS: branch-wise Gaussians for large-scale scenes","Gaussian splatting gets smarter across large view gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000302,"raw_usage":{"total_tokens":1734,"prompt_tokens":934,"completion_tokens":800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":701}},"tokens_in":550,"tokens_out":800,"duration_ms":7145,"temperature":1.0,"reasoning_tokens":701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:45.586634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to train a cross-view scene twice: once with the stated supplementation and once after applying a small rigid perturbation (for example a 10 cm translation) to the branch model coordinates before the voxel-overlap test of Eq. (6), while keeping the rest of the pipeline identical. If the perturbed version matches or exceeds the PSNR of the unperturbed version on held-out aerial or ground views, then the precise voxel alignment is not what drives the gains, and the supplementation mechanism would need a different explanation.","supporting_citations":[{"cited_title":"Scaffold- gs: Structured 3d gaussians for view-adaptive rendering,","cited_arxiv_id":null,"evidence_quote":"Backbone model used for all branch sub-models and the cross-view model; its densification rule is the one being modified."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Defines the original 3DGS adaptive densification control whose gradient averaging the paper identifies as the source of cross-view smoothing."},{"cited_title":"Mip-splatting: Alias-free 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Comparison baseline whose similar-to-3DGS results the paper uses to argue the cross-view problem is not a multi-scale sampling issue."},{"cited_title":"Pygs: Large-scale scene representation with pyramidal 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Motivates the initialization step where the distant-view sub-model is downsampled into a point cloud for the cross-view model."},{"cited_title":"Facenet: A unified em- bedding for face recognition and clustering,","cited_arxiv_id":null,"evidence_quote":"Provides the max(0, ·) triplet-style formulation adopted for the gradient-aware regularization loss."},{"cited_title":"Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond,","cited_arxiv_id":null,"evidence_quote":"Supplies the MatrixCity Block small and Block A datasets used for aerial-ground and pure-ground evaluation."},{"cited_title":"Isprs benchmark for multi-platform photogrammetry,","cited_arxiv_id":null,"evidence_quote":"Supplies the real Zeche scene used to test the method on actual captured aerial-ground data."}],"review_version":1}