{"id":"e7f25e7a-4d35-4b0e-bf7d-2c5971d1b34c","arxiv_id":"2506.13516","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SMW-GS improves unconstrained and large-scale 3D reconstruction by combining micro-macro frustum sampling, wavelet-based frequency features, and visibility-aware camera partitioning.","lead":"SMW-GS is a 3D Gaussian Splatting method for reconstructing scenes from unconstrained photo collections. It separates each point's appearance into global, refined, and intrinsic parts, samples multi-scale features with wavelets and conical frustums, and adds a visibility-aware camera partitioning scheme for large urban scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PSG's \"provable supervision lower bound\" is not guaranteed: Stage 1 can stop on \"no performance gains\" before N_vis reaches tau, and points with fewer than tau visible cameras cannot reach it. This weakens the scalability and boundary-consistency claims.","rationale":"The paper's central claim has two parts: reconstruction quality and scalability. The quality part is supported by extensive comparisons, but the scalability part rests on Point-Statistics-Guided partitioning giving every Gaussian enough supervision. The reader flagged the \"provable lower bound\" as unproven; I agree and sharpen the point: the algorithm's own termination condition (\"no performance gains\") and the existence of points with total visibility below tau mean the bound is not merely unproven but false as a general statement. This is the most load-bearing concern because it attacks the mechanism that is supposed to justify the large-scale gains in Tab. IX and Fig. 12. A direct instrumentation of the released code can settle it: if a nonzero fraction of points remains below tau, the paper must either supply a corrected guarantee (e.g., tau no larger than the minimum visibility) or soften the claim; the empirical SOTA might survive, but the scalability story as written would not. This does not move the verdict, since CONDITIONAL already captures the needed revision.","tokens_in":22774,"tokens_out":7380,"duration_ms":74961,"concrete_test":"Using the released code, instrument Stage 1 to log, after greedy termination for each block in Rubble, Building, Residence, and Sci-Art: (i) the distribution of final N_vis(p_k) against tau = kappa * c_bar; (ii) the fraction of points with N_vis(p_k) < tau, and separately with N_vis(p_k) < 2; (iii) for each block, whether the loop exited because all points met tau or because of \"no performance gains.\" If any nonzero fraction of points remains below tau, the \"provable supervision lower bound\" is false on these scenes, and the Fig. 12 interpretation must be revised; if the fraction is zero, the guarantee holds empirically and the concern is largely theoretical.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. IV-D (Stage 1), the greedy loop terminates when \"forall p_k in B_m, N_vis(p_k) >= tau or no performance gains.\" The paper then states the combined strategy \"ensures provable supervision lower bounds through Stage 1's tau-enforced assignment.\" This does not follow. First, for any point whose total visible camera set has size |V(p_k)| < tau, assigning all cameras observing it still leaves N_vis(p_k) < tau; since tau = kappa * c_bar with kappa in (0,1), sparse or outlier points in real COLMAP clouds will typically have visibility below the average and hence below tau. Second, when the loop stops due to \"no performance gains,\" there is no definition of that condition and no proof that it can only occur once all points are above tau; the algorithm is a greedy heuristic with no stated monotonicity guarantee toward the threshold. The claimed lower bound is therefore not just unproven but false as a general guarantee. Because the boundary-consistency results (Fig. 7, Fig. 12) and the scalability advantage (Tab. IX) are explicitly attributed to this supervision guarantee, the central scalability claim rests on an unsupported and likely violated coverage assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents SMW-GS, a 3D Gaussian Splatting method for unconstrained scene reconstruction. The method decomposes per-Gaussian appearance into global, refined, and intrinsic components; introduces micro-macro projection with jittered conical frustums and wavelet-based sampling; fuses these signals through a Hierarchical Residual Fusion Network; and scales to large scenes via Point-Statistics-Guided camera partitioning and rotational block training. The paper reports consistent PSNR/SSIM/LPIPS gains over prior in-the-wild and large-scale baselines on three Phototourism scenes, four real large-scale scenes (Mill-19 and UrbanScene3D), and four MatrixCity blocks, with particularly large gains on the synthetic appearance-variation blocks. It also provides ablations of the decoupling module, wavelet dimension, sampling count, and partitioning strategy.","tokens_in":23062,"tokens_out":6842,"duration_ms":66859,"significance":"If the reported results are robust, this is a solid contribution to unconstrained and large-scale 3D reconstruction: it extends Gaussian-level appearance disentanglement into divide-and-conquer training, introduces a multi-scale wavelet-based sampling mechanism, and contributes a new synthetic benchmark with multiple appearance conditions. The experimental gains are large and consistent across several datasets, which supports the core claim that the method is competitive. However, the 'provable supervision lower bound' claim is not actually established, and several hyperparameters are tuned on the same test sets used in the main tables, so the significance as stated exceeds what the current evidence demonstrates.","major_comments":[{"comment":"The sentence 'The combined strategy ensures provable supervision lower bounds through Stage 1's τ-enforced assignment' is not supported by the algorithm description. For a point p_k with total visible camera set satisfying |V(p_k)| < τ, the compensation mechanism assigns every camera in V(p_k), leaving N_vis(p_k) = |V(p_k)| < τ; since τ = κ c̄ with κ ∈ (0,1), such points are not excluded by construction. The greedy loop can also stop at 'no performance gains' with no formal definition of that condition, so there is no proof that the loop only terminates after all points reach τ. Because the boundary-consistency gains (Fig. 7, Fig. 12) and the scalability comparison (Tab. IX) are explicitly attributed to this supervision guarantee, the authors should either prove a valid guarantee (e.g., every Gaussian receives all of its visible cameras) or reframe the scheme as a heuristic and report the empirical distribution of N_vis(p_k) after Stage 1.","section":"Sec. IV-D (Stage 1)"},{"comment":"The choices of wavelet level M, frustum sample count k_s, and block grid are made on the same test sets that are reported in the main results. Tabs. VII and VIII tune M and k_s on Brandenburg Gate, Sacre Coeur, and Trevi Fountain, which are exactly the three scenes in Tab. I; Tab. IX selects the {3,2} grid using Rubble and MatrixCity blocks A*/E*, which also appear in Tabs. III and V. This test-set selection makes the reported improvements optimistic and weakens the claim that the method 'consistently outperforms' across datasets. I recommend either fixing hyperparameters a priori, tuning on a separate validation split, or showing that nearby hyperparameter choices give similar results.","section":"Sec. V-D, Tabs. VII–IX"},{"comment":"All quantitative tables report single runs without error bars or significance tests. On Sci-Art the LPIPS value (0.147) is identical to Momentum-GS and the SSIM gap is only 0.005, so the text's claim of 'significantly outperforming' the previous best is not statistically supported for this scene. Adding repeated runs with standard deviations or at least documenting the number of seeds would make the central claim testable.","section":"Sec. V-A, Tables III and V"}],"minor_comments":[{"comment":"The statement that 'its training and rendering remain slower than state-of-the-art methods' appears to contradict Sec. IV-D's 'real-time' claim and Tab. II's 'nearly 1.5 times faster' statement; please clarify the regime (classical vs. large-scale, training vs. rendering).","section":"Sec. VI"},{"comment":"The values of the frustum radii r_dot and R_dot_max, the number of block-grid cells M×N used in the main experiments, and the rotation interval N_iter are not reported; please add them for reproducibility.","section":"Sec. V-B"},{"comment":"The baseline name 'WildGaussian' should be 'WildGaussians' for consistency with the text.","section":"Table III"},{"comment":"The right panel lacks a colorbar and a description of how 'normalized per-point increase' is computed, making the claim about boundary supervision difficult to verify.","section":"Fig. 12"},{"comment":"The claim of 'the first integration of frequency domain data into 3DGS appearance representation' is strong; please qualify it (e.g., 'to our knowledge') and cite any prior wavelet-based 3DGS works if they exist.","section":"Related work"},{"comment":"The abstract says 'optimally assigns camera views'; since Stage 1 is a greedy heuristic, please replace 'optimally' with a more precise description.","section":"Sec. IV-D"}],"recommendation":"major_revision","confidential_remarks":"The core method is plausible and the experiments are extensive, but the 'provable supervision lower bound' is overstated and the hyperparameter selection is partly test-set dependent. I would be willing to review a revised version that fixes the guarantee or reframes it empirically and adds validation-style hyperparameter analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a strong empirical paper with large, consistent gains, and a real new mechanism in the micro-macro wavelet sampling. But the 'provable supervision lower bound' claim in Sec IV-D doesn't hold up, and the paper leans on it for the boundary-consistency results.\n\nWhat's new: the triple decomposition (global/refined/intrinsic) and the Micro-macro Projection with jittered frustums plus wavelet sampling is a real extension of GS-W. The PSG camera partitioning with visibility-aware greedy assignment plus Stage-2 SSIM-based augmentation is a sensible way to adapt Gaussian-level appearance modeling to block-based training. The new MatrixCity renderings with eight appearance conditions are a useful test bed. The authors also disclose the relation to their AAAI MW-GS paper, which is the right thing to do.\n\nWhat works: the empirical results are impressive. Gains of 2.16-3.55 dB PSNR over Momentum-GS on four real large-scale scenes and 6.5/6.4 dB on the two synthetic appearance-variation blocks are large. Ablations in Tab VI and IX show each component contributes. Code is promised. That's real evidence.\n\nWhere it's soft: the PSG guarantee. Stage 1 stops when all points have N_vis >= tau OR 'no performance gains'. There is no definition of 'no performance gains' and no proof the loop reaches tau. Worse, any point whose total visible camera count is below tau can never be brought to tau by assigning all its cameras. So the sentence in Sec IV-D that the combined strategy 'ensures provable supervision lower bounds through Stage 1's tau-enforced assignment' is false as a general claim. The boundary-consistency improvements in Fig 7 and Fig 12 are attributed to exactly this, so the strongest scalability evidence rests on an unearned guarantee. This is fixable: either prove a weaker monotonicity or coverage claim, or drop 'provable' and state explicitly that the threshold is a heuristic target.\n\nAlso, no error bars or significance tests, and hyperparameters M and k_s are selected on the same test sets that appear in the main tables. In this subfield that's common, but it's still worth asking for at least a couple of seeds. The citation pattern looks fine.\n\nWho it's for: anyone working on unconstrained or large-scale 3D reconstruction. It's a strong baseline and the empirical comparison is useful.\n\nRecommendation: this deserves a serious referee, but with a request to fix the guarantee statement and add basic statistical reporting. I'd take it with major revisions, not reject.","headline":"A strong, empirically impressive 3DGS-in-the-wild paper with a real scalability story, undercut by an overclaimed 'provable' supervision guarantee that the algorithm doesn't actually deliver.","tokens_in":23600,"tokens_out":2834,"would_cite":true,"duration_ms":29395,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SMW-GS reports up to 3.55 dB PSNR gains on large-scale unconstrained scenes by splitting appearance into global, refined, and intrinsic components with supervision-aware camera partitioning.","keywords":["3D Gaussian Splatting","unconstrained photo collections","appearance disentanglement","wavelet sampling","large-scale scene reconstruction","camera partitioning","novel view synthesis"],"falsifier":"Run Stage 1 on a low-density block, record the visible-camera count for every Gaussian point at termination, and count how many points still fall below the threshold $\\tau$; any nonzero count would refute the claimed supervision lower bound and the explanation for the reported boundary-region improvements.","tokens_in":22545,"feed_emoji":"🌆","tokens_out":12187,"duration_ms":97695,"temperature":0.7,"pith_summary":"The paper claims that 3D reconstruction from unconstrained photo collections, where lighting and appearance change between shots, can be made both more accurate and more scalable by separating each scene point's appearance into three parts: a global tone, a refined local texture, and an intrinsic material feature. It introduces SMW-GS, a Gaussian-splatting method that samples these features through narrow and broad conical frustums and through wavelet-decomposed frequency bands, then fuses them with a hierarchical residual network. For large scenes, a camera-partitioning strategy assigns views to blocks by per-point visibility so that every Gaussian receives enough supervision. On four real large-scale scenes the method reports PSNR gains of 1.22 to 3.55 dB over the previous best large-scale baseline, and larger gains on synthetic scenes with strong appearance variation.","feed_headline":"Gaussian splatting gains up to 3.55 dB on urban scenes","feed_subtitle":"Splitting appearance into global, refined, and intrinsic parts keeps large unconstrained scenes sharp and consistent.","key_machinery":"The load-bearing machinery is Micro-macro Wavelet-based Sampling (MWS), which projects each 3D Gaussian into a narrow conical frustum, with jittered micro offsets for fine textures, and into a broader frustum, for regional effects such as highlights; it applies a one-level discrete wavelet transform to the feature map so sampling touches low- and high-frequency bands at multiple resolutions. The sampled features are fused by a Hierarchical Residual Fusion Network (HRFN). For scalability, Point-Statistics-Guided (PSG) Camera Partitioning computes how many training cameras see each Gaussian, sets a supervision threshold $\\tau = \\kappa \\bar{c}$, and greedily assigns cameras to blocks until that threshold is met, then adds content-critical cameras detected by an SSIM-difference test. Rotational Block Training alternates blocks across GPUs so the shared appearance network sees the whole dataset. These mechanisms carry the argument by turning per-point appearance variation into per-Gaussian multi-scale sampling, and by converting large-scene partitioning from a geometric heuristic into a supervision-driven assignment.","core_discovery":"The central claim is that appearance variation in unconstrained scenes—lighting changes, shadows, highlights, tone shifts—is better modeled by decomposing each Gaussian's color prediction into global, refined, and intrinsic components, where the refined component comes from sampling the reference image's feature map at both micro and macro scales and in the wavelet domain. The paper further claims that this per-Gaussian appearance disentanglement can be carried into the divide-and-conquer setting of large-scale reconstruction: a Point-Statistics-Guided camera partition plus rotational block training keeps every Gaussian sufficiently supervised, and the result is sharper, more consistent rendering at block boundaries and across the whole scene. Reported evidence includes consistent PSNR, SSIM, and LPIPS improvements over in-the-wild and large-scale baselines, with the largest margins on large urban scenes and on synthetic scenes rendered under eight different appearance conditions.","pith_inferences":["Beyond the paper, the same per-point visibility count could drive camera assignment in other block-based radiance-field or splatting pipelines, since it only needs a reconstructed point cloud and camera poses.","Beyond the paper, the wavelet stage suggests a testable extension: deeper or learned wavelet decompositions might improve detail further, though the paper's $M=1$ result hints at a plateau on classical scenes.","Beyond the paper, the synthetic benchmark with eight appearance conditions offers a controlled arena where future appearance-disentanglement methods can be compared independently of real-data noise.","Beyond the paper, the demonstrated day-to-dusk manipulation implies that region-specific appearance editing could be automated with semantic maps, a step the paper does not take."],"forward_implications":["Large urban scenes with changing daylight or weather reconstruct at higher fidelity than earlier Gaussian-splatting methods, with the largest reported gains on scenes where appearance varies most.","Block-boundary artifacts diminish because boundary Gaussians receive visible cameras from neighboring blocks, as shown in the qualitative boundary comparisons.","Appearance transfer and scene-level relighting become practical: swapping the global feature while keeping intrinsic features changes lighting without retraining geometry.","Rendering stays real-time and storage shrinks, since appearance is compressed into the disentanglement network rather than stored per Gaussian.","The method inherits the real-time rendering and compact storage of its anchor-based Gaussian backbone while adding multi-scale appearance sampling."],"supporting_citations":[{"why":"Supplies the 3D Gaussian representation and tile-based rasterization that the method builds on.","marker":"[5]"},{"why":"Provides the voxel-anchor organization and volume regularization that structure the per-anchor Gaussians.","marker":"[6]"},{"why":"Introduces per-image appearance embeddings, the problem framing SMW-GS extends from unconstrained photo collections.","marker":"[12]"},{"why":"Uses adaptive sampling from projected 2D feature maps; Micro-Macro Projection directly extends this sampling and compares against it.","marker":"[16]"},{"why":"Supplies the Mill-19 large-scale scenes (Rubble, Building) used in the real-world evaluation.","marker":"[49]"},{"why":"A large-scale divide-and-conquer baseline with decoupled appearance modeling that SMW-GS is compared against.","marker":"[52]"},{"why":"A large-scale partitioning baseline whose camera partitioning is replaced by PSG in the 'w/o PSG' ablation.","marker":"[53]"},{"why":"The previous best large-scale baseline that SMW-GS reports beating by 1.22 to 3.55 dB PSNR on real scenes.","marker":"[55]"},{"why":"Supplies the UrbanScene3D scenes (Sci-Art, Residence) used in the large-scale evaluation.","marker":"[67]"},{"why":"Provides the MatrixCity synthetic city environment that is extended with eight appearance conditions for the new benchmark.","marker":"[68]"}],"fun_headline_variants":["Wavelet-based Gaussian split sharpens large scenes","Micro-macro sampling scales urban 3D reconstruction","SMW-GS: decompose to conquer appearance variation","Frequency-domain Gaussians improve scene clarity","Scalable splatting with wavelet color decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Stage-1 greedy camera-assignment loop really gives every Gaussian point the promised minimum number of visible cameras when it stops, since the loop terminates on 'no performance gains' rather than on a check that each point has reached the threshold.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet-based Gaussian split sharpens large scenes","Micro-macro sampling scales urban 3D reconstruction","SMW-GS: decompose to conquer appearance variation","Frequency-domain Gaussians improve scene clarity","Scalable splatting with wavelet color decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000548,"raw_usage":{"total_tokens":2598,"prompt_tokens":906,"completion_tokens":1692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":1621}},"tokens_in":522,"tokens_out":1692,"duration_ms":13462,"temperature":1.0,"reasoning_tokens":1621,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:58:47.483761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Stage 1 on a low-density block, record the visible-camera count for every Gaussian point at termination, and count how many points still fall below the threshold $\\tau$; any nonzero count would refute the claimed supervision lower bound and the explanation for the reported boundary-region improvements.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian representation and tile-based rasterization that the method builds on."},{"cited_title":"Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,","cited_arxiv_id":null,"evidence_quote":"Supplies the Mill-19 large-scale scenes (Rubble, Building) used in the real-world evaluation."},{"cited_title":"Capturing, reconstructing, and simulating: the urbanscene3d dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the UrbanScene3D scenes (Sci-Art, Residence) used in the large-scale evaluation."}],"review_version":2}