{"id":"dfe36b65-c487-46cd-a81a-6d33e3e5b6b5","arxiv_id":"2412.03428","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"2DGS-Room guides 2D Gaussian splats with seed points, monocular depth/normal priors, and multi-view consistency, achieving state-of-the-art indoor reconstruction F-scores.","lead":"A new method builds clean 3D models of indoor rooms from many photos by guiding flat Gaussian splats with seed points and AI depth and normal hints. It reports better surface reconstruction scores on ScanNet and ScanNet++ than previous Gaussian and NeRF approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim omits GaussianRoom [49], the same group's closest indoor Gaussian baseline, from Table 1; a head-to-head comparison on identical splits would settle whether the reported F-scores are actually state-of-the-art.","rationale":"I agree with the reader's overall CONDITIONAL verdict, but I do not think the homography in Eq. 14 is the single most load-bearing weakness. The text says the forward/backward projection is applied 'for each pixel,' so N_r and D_r are plausibly per-pixel rendered depth and normal values rather than a single global plane; and the ablation shows the multi-view loss is a minor contributor (F-score 0.575 vs 0.566). Even if the local-plane homography is inaccurate at occlusions and depth discontinuities, the paper's SOTA claim would not be overturned. The load-bearing premise is instead the comparison set. A state-of-the-art claim is meaningful only relative to the strongest existing methods, and the paper itself identifies GaussianRoom as a closely related indoor Gaussian approach with overlapping authorship, yet excludes it from Table 1. The same can be said for MonoSDF and NeuRIS, which are cited as the protocol references but not compared. Without these comparisons and without released code or per-scene statistics, the abstract's claim is not established. The proposed head-to-head test is feasible because the relevant implementations exist and the authors presumably have access to both systems. Pending that test, conditional acceptance is appropriate; if GaussianRoom outperforms 2DGS-Room, the claim should be revised or rejected.","tokens_in":13968,"tokens_out":6192,"duration_ms":64851,"concrete_test":"Run the released GaussianRoom implementation (or retrain it) on the exact same 8 ScanNet and 4 ScanNet++ scene splits, extract meshes with the same rendered-depth TSDF fusion, and compute F-score with the same threshold. If GaussianRoom's F-score reaches or exceeds 0.575 on ScanNet or 0.464 on ScanNet++, the abstract's SOTA claim fails. If GaussianRoom is clearly below those numbers, the omission is not decisive and the central claim survives this check. Optionally repeat with MonoSDF and NeuRIS under the same protocol to cover neural-implicit baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is comparative: 'state-of-the-art performance in indoor scene reconstruction.' That claim is only as strong as the comparison set, and the comparison set has a conspicuous gap. Section 2.3 cites GaussianRoom [49], a directly related indoor Gaussian method that combines 3DGS with SDF guidance and monocular cues, and the author lists of the two papers overlap. Despite this, Table 1 compares only 3DGS, SuGaR, RaDe-GS, PGSR, and 2DGS among Gaussian methods; GaussianRoom, DN-Splatter, and GSDF are cited but never evaluated. Similarly, the leading neural-implicit indoor baselines MonoSDF [6] and NeuRIS [5] are mentioned as the basis for the evaluation protocol but are absent from Table 1. No code is released, no per-scene results are given, and no error bars are reported, so the headline F-scores (0.575 on ScanNet, 0.464 on ScanNet++) cannot be independently verified. This omission is more load-bearing than the multi-view homography concern in Sec. 4.3: the ablation in Table 2 shows that removing the multi-view loss changes the F-score only from 0.575 to 0.566, so even if Eq. 14 is a poor model at occlusion boundaries, the main result would survive. If GaussianRoom or another omitted indoor method achieves a higher F-score under the same protocol, the SOTA claim is directly falsified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"2DGS-Room modifies 2D Gaussian Splatting for indoor surface reconstruction. The method (i) initializes 2D Gaussians from voxelized, outlier-filtered SfM points (\"seed points\") with learnable offsets, and adjusts seed density during training through gradient-guided growth and opacity-based pruning; (ii) supervises rendered depth and normals with monocular priors (a scale-and-shift-invariant depth loss against Depth Pro predictions, plus L1 and cosine normal losses); and (iii) adds a multi-view loss that warps reference pixels via the plane-induced homography in Eq. (14) and penalizes forward-backward reprojection error (L_geo) and normalized-cross-correlation mismatch (L_pho) against a neighboring view. Training runs for 30k iterations, and meshes are extracted by TSDF fusion of rendered depth maps. On 8 ScanNet and 4 ScanNet++ scenes, the paper reports F-scores of 0.575 and 0.464, the best among the compared methods (NeuS, Neuralangelo, 3DGS, SuGaR, 2DGS, PGSR, RaDe-GS), and the ablation in Table 2 attributes the gains mainly to the seed guidance, depth prior, and normal prior.","tokens_in":14290,"tokens_out":8167,"duration_ms":73866,"significance":"The paper addresses a real gap: 2DGS, while fast, produces floaters and incomplete geometry in indoor scenes, and the reported gains over the 2DGS baseline on identical benchmarks are large (ScanNet F-score 0.341 to 0.575). The three-way ablation in Table 2 is clean and gives explicit credit to each component, and the seed-guided initialization with learned offsets and adaptive density control is plausible and clearly described. The depth and normal priors are standard but sensibly combined, and evaluation uses the standard ScanNet/ScanNet++ protocols with five geometry metrics. However, the headline \"state-of-the-art\" claim is comparative and rests on a comparison set (Table 1) that excludes several directly related methods the paper itself cites, so the significance claim is not yet substantiated; and the multi-view consistency loss rests on a homography whose pixel-wise validity is never derived. Neither issue invalidates the method as presented, but both need to be resolved before the significance claim is supported. Reproducibility is also limited by the absence of code, per-scene numbers, and timing measurements.","major_comments":[{"comment":"The abstract's claim of 'state-of-the-art performance in indoor scene reconstruction' is a comparative claim, and Table 1 does not support it with the strongest relevant comparisons. Section 2.3 cites GaussianRoom [49], DN-Splatter [48], and GSDF [50] as indoor or geometry-aware Gaussian methods, and the evaluation protocol is said to follow MonoSDF [6] and NeuRIS [5]; none of these five methods appears in Table 1. Because GaussianRoom shares authors with this paper and combines 3DGS with SDF guidance and monocular cues, the same ingredients used here, its omission is particularly consequential: a head-to-head comparison on identical scene splits is the most direct way to verify or falsify the stated SOTA claim. Please add these baselines (or give a principled reason for excluding them), report per-scene results for all methods, and soften the abstract claim if the comparison cannot be completed.","section":"§2.3, §5.1, Table 1"},{"comment":"The geometric consistency loss is not derived. As written, H_rn in Eq. (14) uses quantities N_r and D_r that are not defined as per-pixel maps, so the formula is the homography induced by a single global plane; applied pixel-wise to arbitrary indoor geometry it is exact only for pixels lying on that plane and will penalize correct correspondences at depth discontinuities, occlusions, and curved surfaces. If N_r(p) and D_r(p) are per-pixel values taken from the rendered normal and depth maps, the paper must state this and derive the mapping explicitly, and the exclusion set V_e in Eq. (15) must be specified operationally. Note also that the Table 2 ablation limits both the risk and the claimed benefit of this module: removing the multi-view loss changes the F-score from 0.575 to 0.566 and leaves Accuracy and Completion unchanged at 0.055 and 0.092, so the statement that multi-view constraints 'further enhance reconstruction quality' should be calibrated accordingly.","section":"§4.3, Eqs. (14)-(15)"}],"minor_comments":[{"comment":"Please fix the typos: 'Constrains' in the title and abstract, 'start-of-the-art' in Section 1, 'Expect for the center position' in Section 4.1, and 'V oxel-based' in Section 2.1.","section":"Title, Abstract, §1, §2.1, §4.1"},{"comment":"The claim that training is 'approximately 30 times faster' than NeuS/Neuralangelo (over 20 hours) is not supported by any reported wall-clock times; please add a timing table on a common hardware configuration.","section":"§5.2"},{"comment":"Equation (6) is typeset without visible division operators and the phrase 'By homography' is unclear; the formulas for u(x) and v(x) should be restated cleanly so that the ray-splat intersection derivation can be checked.","section":"§3, Eq. (6)"},{"comment":"Table 1 reports only dataset-averaged numbers; please report per-scene results and error bars or variance across scenes, since the F-score differences among methods are the load-bearing evidence for the SOTA claim.","section":"Table 1"},{"comment":"The sentence about methods that 'achieve higher Accuracy than 2DGS on the ScanNet++ dataset' is confusing because Accuracy is a distance metric where lower values are better; please rephrase to clarify that NeuS and SuGaR obtain better (lower) Accuracy values.","section":"§5.2"},{"comment":"The supplement states that the multi-view consistency constraint is applied only after 7,000 iterations, but the main-text optimization section does not mention this schedule; please move this detail into the main text so that the total loss in Eq. (18) is fully specified.","section":"§4.4, Supplementary B"},{"comment":"Reference [13] is formatted incorrectly with 'Robust Multiview Stereopsis' appearing as the author name; the citation should be corrected to the actual author list and title.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The author lists of this manuscript and of GaussianRoom [49] overlap substantially (four co-authors), and GaussianRoom is the closest published indoor Gaussian baseline to the present method. Its absence from Table 1, in a paper whose abstract leads with a state-of-the-art claim, is a comparison-practice issue the editor may wish to raise explicitly with the authors, beyond the scientific request for a head-to-head comparison. The submitted version also contains no code and no per-scene statistics, which further limits independent verification of the headline F-scores."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The seed-guided mechanism is the real addition here, and the gains over 2DGS on ScanNet are large enough to take seriously. Voxelizing SfM points into seeds, attaching learnable offsets, and then growing/pruning seeds by gradient and opacity is a sensible way to steer Gaussians toward surfaces, and the ablation shows it carries most of the improvement (F-score drops from 0.575 to 0.307 without it). The monocular depth and normal priors are standard, but they are integrated cleanly and the ablations suggest each helps. I also credit the paper for reporting component-wise ablations at all; that is more than many GS papers do.\n\nThe soft spots are real but not all equally soft. The multi-view homography in Eq. 14 assumes a single plane per reference view, which is obviously wrong for indoor scenes with occlusions and non-planar objects. But the ablation shows removing the multi-view loss changes F-score only from 0.575 to 0.566, so this is a minor flaw, not a load-bearing one. What actually undermines the SOTA claim is the comparison set. GaussianRoom [49] is cited as related work, the author lists overlap with this paper, and it is arguably the closest indoor-GS baseline, yet it is absent from Table 1. DN-Splatter and GSDF are also omitted. The SOTA claim is only as good as the baselines you choose, and this choice is conspicuous. No code, no per-scene results, and no error bars make it worse: the headline numbers cannot be independently checked, and the 0.575 could easily be within noise of a well-tuned GaussianRoom run on the same scenes.\n\nSo where does this leave the paper? The core pipeline is reasonable, the seed-guided idea is novel enough, and the evaluation protocol follows standard practice. But the missing baseline is not a nit; it is central to the advertised contribution. A serious referee should ask for a head-to-head with GaussianRoom (and ideally DN-Splatter), per-scene numbers, and either code or a clear statement of what was rerun. I would not cite the F-scores as established until that happens, but I would read a revised version.\n\nFor peer review: send it out. The topic is active, the method is a plausible engineering contribution, and the issues are fixable in revision. A desk reject would be premature; a hard accept with the current evidence would be too generous. Major revision.","headline":"A plausible seed-guided 2DGS extension with solid ablations, but the SOTA claim needs GaussianRoom in the table before I would trust it.","tokens_in":14849,"tokens_out":2022,"would_cite":false,"duration_ms":21626,"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":"2DGS-Room introduces seed-guided 2D Gaussian splatting for indoor scene reconstruction, claiming state-of-the-art geometric fidelity on ScanNet and ScanNet++ benchmarks.","keywords":["indoor scene reconstruction","2D Gaussian splatting","seed point guidance","monocular depth prior","monocular normal prior","multi-view consistency","surface reconstruction","3D reconstruction"],"falsifier":"Measure the per-pixel forward-backward projection error from Eq. (14)-(15) on a scene with strong occlusions and depth discontinuities; if a substantial fraction of pixels (e.g., over 30%) are excluded by the valid-pixel mask, then the geometric consistency loss is only applied where the planar assumption holds, and its contribution would be negligible or harmful on non-planar scenes. A direct comparison of F-score gains from the multi-view term on a near-planar scene versus a heavily occluded scene would reveal this dependence.","tokens_in":13736,"feed_emoji":"🏠","tokens_out":3102,"duration_ms":29564,"temperature":0.7,"pith_summary":"This paper attempts to show that indoor scene reconstruction can be made both accurate and fast by replacing neural radiance fields with a structured 2D Gaussian splatting approach. The proposed method, 2DGS-Room, controls where 2D Gaussian primitives are placed and how densely they are distributed using seed points derived from structure-from-motion, then refines geometry with monocular depth and normal priors plus multi-view consistency constraints. On the ScanNet and ScanNet++ datasets, the authors report F-scores of 0.575 and 0.464, the best among compared methods, implying that explicit splat-based surfaces can rival SDF-based neural methods in fidelity while training about 30 times faster. If correct, this gives a practical route to high-quality 3D reconstruction of rooms from ordinary photo collections.","feed_headline":"Seed-guided 2D Gaussians beat prior methods on indoor meshes","feed_subtitle":"Monocular priors and multi-view checks lift room reconstruction fidelity on ScanNet and ScanNet++.","key_machinery":"The central mechanism is a seed-guided 2D Gaussian representation. Seed points are generated by voxelizing a filtered SfM point cloud, and each seed spawns a set of 2D Gaussian splats whose positions are learnable offsets from the seed. An adaptive growth and pruning strategy adjusts seed density based on gradient magnitude and cumulative opacity. Geometry is further constrained by scale-shift-invariant depth supervision from a monocular depth predictor and L1 plus cosine-similarity normal supervision from a monocular normal estimator. Multi-view consistency is enforced through a homography-based geometric loss and a normalized cross-correlation photometric loss between neighboring views, with a final loss combining these terms.","core_discovery":"The central claim is that seed-guided initialization and optimization of 2D Gaussian splats, combined with monocular geometric priors and multi-view geometric and photometric consistency, yields indoor surface reconstructions with higher fidelity than existing Gaussian splatting and NeRF baselines. The paper quantifies this with F-scores of 0.575 on ScanNet and 0.464 on ScanNet++, outperforming methods such as 2DGS, SuGaR, PGSR, and NeuS on the same datasets. The contribution is an extension of 2DGS that addresses its weaknesses in indoor scenes: scattered floaters, missed textureless regions, and misaligned surfaces.","pith_inferences":["The homography-based geometric consistency loss assumes a single plane per reference image; in non-planar scenes with occlusions and depth discontinuities, this assumption may cause the loss to penalize correct correspondences unless the pixel mask excludes such regions.","The dependence on pre-trained monocular depth and normal predictors means the method's upper bound is tied to the accuracy of those priors, so improvements in monocular estimation would likely transfer directly to reconstruction gains.","A natural extension would replace the global homography with per-pixel depth-based reprojection using rendered depths, which would handle non-planar geometry more gracefully.","The seed-guided growth and pruning strategy could be adapted to other primitive-based representations or to dynamic scenes, where controlling primitive density spatially is equally important."],"forward_implications":["If the reported results hold, indoor surface reconstruction from RGB-only input can reach high fidelity without SDF-based neural rendering, cutting training time dramatically.","Seed-guided distribution of Gaussian primitives provides a practical way to enforce scene structure, reducing floating artifacts that plague unconstrained Gaussian splatting.","Monocular depth and normal priors compensate for textureless indoor regions, yielding smoother and more complete walls, floors, and ceilings.","Multi-view consistency constraints offer a lightweight refinement that reduces view-dependent inconsistencies in reconstructed meshes.","The method's success suggests that explicit primitive-based representations, not just implicit surfaces, can be the foundation for high-quality 3D reconstruction."],"supporting_citations":[{"why":"Provides the base 2D Gaussian splatting representation and rendering pipeline that this work extends with seed guidance and geometric constraints.","marker":"[10]"},{"why":"Supplies the SfM-derived point cloud that is filtered and voxelized to generate seed points for Gaussian initialization.","marker":"[47]"},{"why":"Generates the monocular depth prior used in the scale-shift-invariant depth supervision loss.","marker":"[54]"},{"why":"Generates the monocular surface normal prior used in the normal supervision loss.","marker":"[56]"},{"why":"Defines the scale-and-shift-invariant depth alignment used to compare rendered and reference depths.","marker":"[55]"},{"why":"Provides the normalized cross-correlation measure used in the photometric consistency loss between views.","marker":"[59]"},{"why":"Supplies the ScanNet indoor scenes used for evaluation and comparison with baselines.","marker":"[57]"},{"why":"Supplies the ScanNet++ indoor scenes used for evaluation and comparison with baselines.","marker":"[58]"}],"fun_headline_variants":["Seeded 2D Gaussians outperform on indoor meshes","Seed-guided 2D Gaussian splatting lifts indoor reconstruction","Indoor mesh fidelity rises with seeded Gaussian splatting","2DGS-Room: seed-guided 2DGS sharpens indoor surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The multi-view geometric consistency loss assumes each reference image can be approximated as a single plane with one depth and one normal, which indoor scenes with occlusions and depth jumps violate.","fun_headline_variants_meta":{"raw":{"variants":["Seeded 2D Gaussians outperform on indoor meshes","Seed-guided 2D Gaussian splatting lifts indoor reconstruction","Indoor mesh fidelity rises with seeded Gaussian splatting","2DGS-Room: seed-guided 2DGS sharpens indoor surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2152,"prompt_tokens":850,"completion_tokens":1302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":1228}},"tokens_in":466,"tokens_out":1302,"duration_ms":10155,"temperature":1.0,"reasoning_tokens":1228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:23:54.549317+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the per-pixel forward-backward projection error from Eq. (14)-(15) on a scene with strong occlusions and depth discontinuities; if a substantial fraction of pixels (e.g., over 30%) are excluded by the valid-pixel mask, then the geometric consistency loss is only applied where the planar assumption holds, and its contribution would be negligible or harmful on non-planar scenes. A direct comparison of F-score gains from the multi-view term on a near-planar scene versus a heavily occluded scene would reveal this dependence.","supporting_citations":[{"cited_title":"Structure- from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Supplies the SfM-derived point cloud that is filtered and voxelized to generate seed points for Gaussian initialization."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer","cited_arxiv_id":null,"evidence_quote":"Defines the scale-and-shift-invariant depth alignment used to compare rendered and reference depths."},{"cited_title":"Fast normalized cross- correlation","cited_arxiv_id":null,"evidence_quote":"Provides the normalized cross-correlation measure used in the photometric consistency loss between views."},{"cited_title":"Scannet: Richly-annotated 3d reconstructions of indoor scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the ScanNet indoor scenes used for evaluation and comparison with baselines."},{"cited_title":"Scannet++: A high-fidelity dataset of 3d in- door scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the ScanNet++ indoor scenes used for evaluation and comparison with baselines."}],"review_version":1}