{"id":"bdb6c46b-77d4-4899-adbe-542d70c6fe5a","arxiv_id":"2412.03451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An optimization method that fits 3D rectangular plane primitives to posed indoor photos via differentiable splatting, reconstructing planar scenes in about three minutes without plane annotations.","lead":"PlanarSplatting reconstructs indoor rooms as flat planes in about three minutes from ordinary photos, without needing labeled plane data. It adjusts thousands of 3D rectangles until their projected depth and surface directions match the images, which is fast enough for practical indoor 3D modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only supervision is monocular depth/normal pseudo-labels (Eq. 14); if Metric3Dv2 or Omnidata were trained on ScanNet/ScanNet++, the reported accuracy advantage may reflect prior overfitting rather than the proposed plane-splatting geometry.","rationale":"The reader's weakest assumption and my load-bearing concern converge: PlanarSplatting's optimization is supervised exclusively by monocular depth and normal pseudo-labels (Eq. 14), so the accuracy ceiling is set by Metric3Dv2 and Omnidata. If those foundation models saw the evaluation datasets during training, the benchmark results are optimistic. I agree with the reader that this warrants a conditional verdict rather than acceptance. I considered other possible concerns, such as the absence of code, lack of error bars, and the discrepancy between 'hundreds of scenes' and the actual 130 evaluated scenes, but these primarily affect confidence and reproducibility rather than the core correctness of the method. The monocular-prior dependence is more load-bearing because it directly affects whether the measured geometric accuracy is real on held-out scenes. My proposed test, evaluating on a disjoint indoor dataset, would settle whether the advantage generalizes. Since the reader already identified this as the weakest assumption and assigned CONDITIONAL, no verdict change is needed.","tokens_in":12017,"tokens_out":7383,"duration_ms":78933,"concrete_test":"Check the training data lists of Metric3Dv2 (arXiv:2404.15506) and Omnidata (ICCV 2021) for inclusion of ScanNet or ScanNet++; then run PlanarSplatting on a held-out indoor dataset such as Replica or Matterport3D using the same evaluation protocol, and compare against AirPlanes and PlanarRecon. If the reported accuracy advantage shrinks or reverses on the held-out scenes, the central claim is dataset-dependent rather than a general property of the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that PlanarSplatting reconstructs indoor scenes in 3 minutes with significantly better geometric accuracy than prior plane reconstruction methods. The optimization in Sec. 3.2–3.3 is driven solely by L_render (Eq. 14), which penalizes differences between rendered depth/normal maps and pseudo-labels from Metric3Dv2 and Omnidata. There is no photometric consistency term, no multi-view geometric consistency term, and no direct supervision from input RGB or 3D ground truth. Consequently the reconstructed planes can be no more accurate than the monocular priors; any systematic bias in those priors, e.g., on textureless walls, curved surfaces, or atypical layouts, is directly baked into the output. This is especially concerning because both foundation models are trained on large indoor scan collections and may include ScanNet/ScanNet++ data. If so, the benchmark comparisons in Tables 1–2 do not test generalization to new scene distributions. Additionally, the reported '3 minutes' likely covers only plane optimization, not Metric3Dv2/Omnidata inference time, and the ablations in Table 3 do not include a weak-prior or no-prior condition, so the contribution of the splatting function relative to the priors is not isolated. The paper explicitly acknowledges curved surfaces as a limitation, but does not address the more fundamental dependence of the entire reconstruction on potentially biased monocular supervision.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PlanarSplatting, an optimization-based method for planar surface reconstruction from posed multi-view images. The scene is represented as a set of learnable 3D rectangular plane primitives, and a differentiable plane splatting function renders depth and normal maps from these primitives. The primitives are optimized against monocular depth and normal pseudo-labels obtained from Metric3Dv2 and Omnidata, without any 2D/3D plane annotations. A custom CUDA implementation is said to reconstruct a scene in about 3 minutes. Experiments on ScanNet and ScanNet++ report that the method outperforms prior geometry-based and plane-annotation-based baselines on Chamfer distance and F-score, and that initializing 3DGS/2DGS with the reconstructed planes improves novel view synthesis quality while reducing training time.","tokens_in":12329,"tokens_out":3654,"duration_ms":33895,"significance":"If the reported results hold, the paper makes a useful contribution to indoor planar reconstruction: it removes the need for plane detection, matching, and tracking, avoids plane-annotation supervision, and achieves a remarkably fast per-scene optimization time. The rectangle-aware plane splatting function and its CUDA implementation are technically interesting. The evaluation is also larger than most prior work, covering 130 scenes across two datasets. However, the evidence as presented is incomplete: the supervision is entirely provided by external monocular priors that may have been trained on the evaluation datasets, no statistical significance or error bars are reported, and the code is not released, so the central 'accurate in 3 minutes' claim cannot currently be independently verified.","major_comments":[{"comment":"The only supervision for the plane optimization is the render loss L_render against pseudo-labels from Metric3Dv2 and Omnidata. If these foundation models were trained on ScanNet or ScanNet++, the reported improvements on those datasets may reflect the priors' familiarity with the test distribution rather than the proposed splatting optimization. The paper should report whether the training data of Metric3Dv2 and Omnidata overlap the evaluation scenes, or add a held-out evaluation on distributions not seen by the priors, and should include an ablation that replaces the specific priors with a different depth/normal source (or a photometric-only loss) to isolate the contribution of the plane splatting optimization.","section":"Sec. 3.2, Eq. (14); Sec. 4.2, Tables 1-2"},{"comment":"No error bars, confidence intervals, or significance tests are reported. On ScanNet, the improvement over AirPlanes is modest (Chamfer 5.30 vs 4.83; F-score 64.92 vs 68.85) and could be within scene-level variance. Please report per-scene standard deviations or perform a paired significance test across the 100 and 30 scenes to support the claim of 'significantly better geometric accuracy'.","section":"Sec. 4.2, Tables 1-2"},{"comment":"The plane merge criterion states that primitives are merged if their normal angle error is below 25 degrees and their offset distance error is below 0.1 cm. An offset threshold of 0.1 cm (1 mm) is implausibly tight for merging planes in scene-scale reconstruction and would likely prevent any meaningful merging; this appears to be a likely typo (e.g., 0.1 m). Please clarify the intended value and report sensitivity of the final metrics to this threshold.","section":"Sec. 3.3, Plane Merge"},{"comment":"The '3 minutes' claim appears to refer only to the 5,000-iteration plane optimization. The wall-clock time also includes Metric3Dv2 depth inference, Omnidata normal inference, and the final plane merging. Please report the full pipeline time, including pseudo-label generation and merging, and specify the GPU hardware used for each component. This is important because the stated speed advantage over baselines may be reduced when prior inference is counted.","section":"Sec. 3.1 and Sec. 4.4"},{"comment":"The manuscript states that the CUDA implementation will be released after publication but does not provide code or a detailed algorithmic description of the CUDA kernels. Given that the method's speed and geometry quality depend critically on the CUDA implementation of the differentiable splatting function, the results are not reproducible from the paper alone. I strongly encourage releasing the code (or a detailed pseudocode with performance benchmarks) with the paper or in the supplementary material.","section":"Sec. 1 and Sec. 4.1"}],"minor_comments":[{"comment":"The abstract claims evaluation 'over hundreds of scenes,' but the experiments use 100 ScanNet scenes and 30 ScanNet++ scenes, totaling 130, which is more than one hundred but not 'hundreds.' Please adjust the wording.","section":"Abstract"},{"comment":"There is a typo: 'deoth' should be 'depth.'","section":"Sec. 3.2, after Eq. (13)"},{"comment":"The metric header 'LIPPS' should be 'LPIPS.'","section":"Table 4"},{"comment":"The splitting criterion uses 'average radii gradients' greater than 0.2, but it is not specified whether this average is taken over the pixels covered by the plane, over the plane's area, or over some other set. Please clarify the exact aggregation used.","section":"Sec. 3.3, Plane Splitting"},{"comment":"For ScanNet++, the paper says 'we randomly select 30 scenes for evaluation'; please provide the random seed or the list of scene ids to make the evaluation reproducible.","section":"Sec. 4.1, Datasets"}],"recommendation":"major_revision","confidential_remarks":"The strongest concern is the potential overlap between the training data of the monocular priors and the ScanNet/ScanNet++ evaluation sets. Even if this is not a deliberate flaw, it weakens the claim of a superior planar reconstruction method. I would like the editor to ask for: (1) a statement on whether Metric3Dv2 and Omnidata were trained on ScanNet or ScanNet++ (or a held-out evaluation), (2) error bars or significance tests, and (3) the CUDA code or at least a detailed algorithmic description. The paper is otherwise technically sound and within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bin, here's my take on PlanarSplatting. The core idea is genuinely new: instead of detecting and matching planes in image space, they directly optimize 3D rectangular primitives with a rectangle-aware splatting function, supervised by monocular depth/normal pseudo-labels. Three minutes per scene and no plane annotations is a real practical step forward, and the numbers on ScanNet and ScanNet++ back up the claim. The ablations show the double radii and the splitting rule each matter, and the sphere-initialization baseline suggests the monocular init mainly buys convergence speed, not final quality. I also like the GS integration: using the planes to initialize 3DGS/2DGS without densification beats the vanilla versions in both quality and time.\n\nThat said, the supervision is the soft underbelly. Everything is driven by Eq. 14, which only compares rendered depth/normal maps to Metric3Dv2 and Omnidata outputs. There is no photometric term and no multi-view consistency, so the planes can only be as good as those priors. If either foundation model was trained on ScanNet or ScanNet++, the benchmark comparisons become optimistic. The paper never mentions this. A proper rebuttal would need a cross-dataset test or at least a no-prior/photometric-only variant. As is, the contribution of the splatting function relative to the priors is not isolated.\n\nOther issues are smaller. The abstract says 'hundreds of scenes', but the evaluation is 100 ScanNet + 30 ScanNet++ scenes. The 3-minute figure excludes the time to run Metric3Dv2 and Omnidata inference, which is not trivial. There are no error bars or significance tests anywhere. The ablation is on 10 scenes. And the CUDA code is not actually available yet, only promised, so I cannot verify the timing or the tables.\n\nNone of this kills the paper. The method is sensible, the writing is clear, and the claimed accuracy is plausible. I would send it to review, and the reviewers should push for code release, error bars, and a careful discussion of training-set overlap for the priors.","headline":"A practical plane-splatting method that delivers on speed, but the monocular-prior supervision leaves a data-overlap confound unexamined.","tokens_in":12871,"tokens_out":2432,"would_cite":true,"duration_ms":23259,"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":"Indoor scenes reconstruct as accurate 3D planes in about 3 minutes, with no plane-labeled training data","keywords":["planar reconstruction","3D plane primitives","differentiable splatting","Gaussian splatting","indoor scenes","multi-view reconstruction","novel view synthesis","monocular priors"],"falsifier":"Reconstruct a held-out set of indoor scenes whose ground-truth geometry comes from a laser scan and that were not used to train the monocular depth and normal priors, then compare plane Chamfer distance and F-score with the baselines reported in Tables 1 and 2. If the margins disappear or reverse, the benchmark numbers are inflated by prior-model exposure rather than by the plane optimization itself.","tokens_in":11804,"feed_emoji":"🏠","tokens_out":10971,"duration_ms":96742,"temperature":0.7,"pith_summary":"PlanarSplatting tries to establish that a full indoor scene can be reconstructed as a compact set of 3D planes by optimizing rectangular plane primitives directly in 3D space, instead of detecting, matching, and tracking planes in images. Rendering these primitives into depth and normal maps, the method supervises itself with monocular depth and normal predictions, so it needs no plane-labeled training data. The paper reports that the optimization converges in about 3 minutes per scene on ScanNet and ScanNet++, with lower Chamfer distance and higher F-score than prior plane-reconstruction methods, including ones trained with plane annotations. The same plane primitives also initialize Gaussian-splatting renderers, improving novel-view synthesis while using fewer points and less training time. If the claim holds, structured indoor geometry becomes cheap enough to generate at scale.","feed_headline":"3-minute reconstruction turns rooms into 3D planes","feed_subtitle":"No plane labels or detection step: rooms become structured 3D planes in 3 minutes.","key_machinery":"The load-bearing object is the rectangular 3D plane primitive, parameterized by a center $p_\\pi$, a rotation quaternion $q_\\pi$, and double-direction radii along its local X and Y axes; the double radii give the patch independent extent on either side of its center. The argument is carried by the differentiable plane splatting function: for a ray-plane intersection, the splatting weight is the minimum of a sigmoid-shaped weight along the X-axis and one along the Y-axis, with a sharpness hyperparameter $\\lambda$ that grows during optimization, so the primitive approximates a true rectangle rather than a soft Gaussian ellipse. These weights are used to composite depth and normal maps over the 30 nearest intersections per ray, and the render loss against monocular depth and normal priors drives all optimization. Plane splitting by radius gradients during optimization, a final merge by normal angle and offset thresholds, and the CUDA forward/backward implementation complete the mechanism that produces accurate plane instances within 3 minutes.","core_discovery":"The central claim is that planar surface reconstruction can be formulated as direct optimization of explicit 3D plane primitives $\\pi = (p_\\pi, q_\\pi, r_x^+, r_x^-, r_y^+, r_y^-)$, rectangular patches with a center, a quaternion rotation, and four directional radii. Each primitive is rendered into a ray's depth and normal through a rectangular splatting weight — the minimum of two sigmoid-shaped weights along the patch's local axes — and the nearest intersections are $\\alpha$-composited front to back. The only supervision comes from comparing these rendered maps with monocular depth and normal predictions, and a radius-gradient-based splitting operation lets oversized primitives adapt to multiple coplanar regions. The paper argues that this removes the need for 2D/3D plane detection and cross-view plane matching, and reports the best geometry metrics among compared methods on both datasets while reconstructing a scene in about 3 minutes thanks to a CUDA implementation.","pith_inferences":["Beyond the paper: because the only supervision is monocular depth and normal prediction, PlanarSplatting inherits any systematic bias of those priors; if the priors were trained on the benchmark datasets themselves, the reported margins over ScanNet and ScanNet++ baselines could be optimistic.","Beyond the paper: the pipeline's speed and lack of plane labels make it a natural pseudo-label generator for training plane detectors or building large structured indoor datasets, a use the paper gestures at in its conclusion.","Beyond the paper: the rectangle-shaped splatting function is a general differentiable-rendering idea; the same shape-aware window could be applied to other parametric primitives such as boxes, cylinders, or superquadrics.","Beyond the paper: the final plane count should be controllable through initialization density and the gradient threshold for splitting, so the representation's level of detail is a tunable knob; this is not reported in the paper."],"forward_implications":["Plane detection, matching, and tracking stages are not needed: multi-view plane consistency is enforced by optimizing 3D primitives directly against rendered depth and normal maps.","On the paper's results, the method achieves lower Chamfer distance and higher F-score than PlanarRecon and AirPlanes on ScanNetV2 and ScanNet++, even though those baselines use plane annotations during training.","Gaussian splatting renderers can be initialized from the reconstructed plane primitives, and with point positions fixed and densification removed, the paper reports better PSNR and SSIM than vanilla 3DGS and 2DGS while using fewer points and less total time.","Because the output is a merged collection of plane primitives, the representation is compact and structurally explicit, suitable for editing or downstream modeling.","The method targets planar indoor scenes; curved surfaces and strongly non-planar geometry are outside its intended scope."],"supporting_citations":[{"why":"Supplies the monocular depth prior that initializes the plane primitives and serves as the depth pseudo-label in Eq. (14).","marker":"[9]"},{"why":"Supplies the monocular surface-normal prior used as the normal pseudo-label in Eq. (14).","marker":"[3]"},{"why":"Provides the ScanNetV2 posed-image benchmark and the test split on which geometry and segmentation metrics are reported.","marker":"[2]"},{"why":"Provides the ScanNet++ posed-image benchmark used for the cross-dataset evaluation.","marker":"[38]"},{"why":"Defines the volumetric plane-tracking paradigm the paper removes, and supplies baseline results and evaluation metrics for planar reconstruction.","marker":"[35]"},{"why":"Supplies the AirPlanes baseline, the RANSAC plane-extraction implementation used for geometry-based baselines, and the planar-fidelity evaluation protocol.","marker":"[33]"},{"why":"Supplies 2D Gaussian splatting as an NVS baseline and as a mesh-plus-RANSAC baseline, and prior differentiable splatting machinery.","marker":"[10]"},{"why":"Supplies 3D Gaussian splatting as the main NVS baseline and the differentiable Gaussian splatting framework the paper builds on.","marker":"[12]"},{"why":"Provides the ground-truth 3D plane annotations used to evaluate plane reconstruction quality on ScanNetV2.","marker":"[17]"}],"fun_headline_variants":["Splat planes, skip detection: rooms in 3D in 3 min","Direct 3D plane splatting: 3-minute rooms, no detection","3D planes from RGB: no matching, just 3-minute splats","Planar reconstruction without detection: done in 3 minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction is only as accurate as the monocular depth and normal priors that initialize the planes and provide the only supervision; if those priors are systematically wrong for a scene, the optimized plane primitives converge to the wrong surface.","fun_headline_variants_meta":{"raw":{"variants":["Splat planes, skip detection: rooms in 3D in 3 min","Direct 3D plane splatting: 3-minute rooms, no detection","3D planes from RGB: no matching, just 3-minute splats","Planar reconstruction without detection: done in 3 minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3286,"prompt_tokens":964,"completion_tokens":2322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2241}},"tokens_in":580,"tokens_out":2322,"duration_ms":14828,"temperature":1.0,"reasoning_tokens":2241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:23:34.554587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a held-out set of indoor scenes whose ground-truth geometry comes from a laser scan and that were not used to train the monocular depth and normal priors, then compare plane Chamfer distance and F-score with the baselines reported in Tables 1 and 2. If the margins disappear or reverse, the benchmark numbers are inflated by prior-model exposure rather than by the plane optimization itself.","supporting_citations":[{"cited_title":"Omnidata: A scalable pipeline for making multi- task mid-level vision datasets from 3d scans","cited_arxiv_id":null,"evidence_quote":"Supplies the monocular surface-normal prior used as the normal pseudo-label in Eq. (14)."},{"cited_title":"Chang, Manolis Savva, Maciej Halber, Thomas A","cited_arxiv_id":null,"evidence_quote":"Provides the ScanNetV2 posed-image benchmark and the test split on which geometry and segmentation metrics are reported."},{"cited_title":"Scannet++: A high-fidelity dataset of 3d indoor scenes","cited_arxiv_id":null,"evidence_quote":"Provides the ScanNet++ posed-image benchmark used for the cross-dataset evaluation."},{"cited_title":"Planarrecon: Realtime 3d plane detection and reconstruction from posed monocular videos","cited_arxiv_id":null,"evidence_quote":"Defines the volumetric plane-tracking paradigm the paper removes, and supplies baseline results and evaluation metrics for planar reconstruction."},{"cited_title":"Brostow, Michael Firman, and Sara Vicente","cited_arxiv_id":null,"evidence_quote":"Supplies the AirPlanes baseline, the RANSAC plane-extraction implementation used for geometry-based baselines, and the planar-fidelity evaluation protocol."},{"cited_title":"2d gaussian splatting for geometrically accurate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies 2D Gaussian splatting as an NVS baseline and as a mesh-plus-RANSAC baseline, and prior differentiable splatting machinery."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies 3D Gaussian splatting as the main NVS baseline and the differentiable Gaussian splatting framework the paper builds on."},{"cited_title":"Planercnn: 3d plane detection and reconstruction from a single image","cited_arxiv_id":null,"evidence_quote":"Provides the ground-truth 3D plane annotations used to evaluate plane reconstruction quality on ScanNetV2."}],"review_version":1}