{"id":"a5d33be2-359a-4bd6-bfb4-8bffe3625ec6","arxiv_id":"2411.19271","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Adaptive filtering of noisy phone depth and monocular normal priors improves Gaussian splatting based 3D indoor reconstruction and mesh extraction.","lead":"This paper presents a method to combine smartphone depth sensor readings with AI-predicted surface normals into Gaussian splatting pipelines for 3D room reconstruction. It filters unreliable depth and normal estimates during training and uses an octree-based meshing step to extract finer surface details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DNC's depth-normal estimator as written uses the maximum PCA eigenvector (a tangent direction), not the surface normal, and Eqs. (4)/(8) set filtered targets to 0; the mechanism driving the claimed gains is not reproducible as described.","rationale":"The reader's weakest_assumption (DNC/ANR arbitration) is reasonable, but the text has a more basic problem: the DNC normal estimator is specified as the maximum PCA eigenvector, which is not a surface normal, and the masking equations are written as zero targets. If the implementation matched the text, DNC would not produce the reported improvements; the only way the results can stand is if the code uses the minimum eigenvector and a binary mask. This is exactly the kind of internal inconsistency that should be resolved before the central claim is taken at face value. The ablation table gives real evidence that the implemented filters help, so I do not move to REJECT; I keep the reader's CONDITIONAL verdict, with the added condition that the DNC description be corrected and the filter decisions be evaluated against ground truth. Appendix E ('does not consistently enhance the overall quality of 3D reconstructions') should also be reflected by softening the meshing claim in the abstract, but the DNC specification issue is the most load-bearing. A quick code inspection will settle the main point.","tokens_in":16612,"tokens_out":13149,"duration_ms":119190,"concrete_test":"Inspect the released code's DNC function: (1) which eigenvector of the KNN covariance is selected as Nd, and (2) whether masked pixels enter LD/LN as zero targets or as dropped terms. Then, on a MuSHRoom scene with ground-truth mesh, compute precision/recall of the 10° DNC mask by checking whether masked depth pixels actually have larger GT depth error than unmasked ones. If the code uses the maximum eigenvector or zero targets, the mechanism as claimed fails; if it uses the minimum eigenvector and a mask, the concern reduces to a description error, and the precision/recall result would validate or refute the arbitration assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 says Nd is 'the maximum eigenvector corresponding to the maximum eigenvalue' of the KNN covariance (k=200). For a locally planar depth patch, the surface normal is the eigenvector of the smallest eigenvalue; the maximum eigenvector lies in the tangent plane. If implemented as written, θd in Eq. (5) would be near 90° on planar regions, and with τd=10° (Supp. A.1) DNC would mask essentially all depth after Td=7k, so DNC could not explain the gains in Table 3. If the code uses the minimum eigenvector, the text is simply wrong. Similarly, Eqs. (4) and (8) write filtered values as 0; under the L1 losses in Eqs. (6)-(7), that would pull rendered depths and normals toward zero at masked pixels instead of dropping the term. The released code must be checked to see whether a mask is applied. This matters because DNC and ANR are the core contribution, and the 10° arbitration between iPhone depth and Omnidata normals has no ground-truth validation. Appendix E also concedes that the IsoOctree mesh does not consistently improve overall mesh quality, so the abstract's meshing claim should be narrowed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AGS-Mesh, a plug-in supervision strategy for Gaussian Splatting (3DGS and 2DGS) that adaptively filters noisy iPhone depth maps and monocular normal priors during optimization, plus a depth-adaptive TSDF/IsoOctree meshing post-process. Depth Normal Consistency (DNC) compares normals derived from local depth patches with Omnidata monocular normals to suppress unreliable depth, and Adaptive Normal Regularization (ANR) suppresses normal supervision where rendered normals disagree with monocular predictions. Experiments on MuSHRoom (6 scenes) and ScanNet++ (2 scenes) report improved mesh reconstruction metrics and novel-view synthesis over several baselines, including DN-Splatter, 2DGS, and Splatfacto. The main contribution is the adaptive filtering scheme; the meshing contribution is a separate post-processing strategy.","tokens_in":16874,"tokens_out":3599,"duration_ms":32848,"significance":"If the filtering behavior is implemented as intended, the method is a practical and potentially useful plug-in for improving geometric quality of Gaussian Splatting reconstructions from noisy smartphone captures, and the release of code is a strength. The paper also shows that adding the regularizers reduces Gaussian count while improving or maintaining rendering quality, which is a nontrivial and useful observation. However, the significance is currently conditional because the written description of the core filtering mechanism contains technical inconsistencies that prevent reproduction, the ScanNet++ evaluation is limited to two scenes with no variance reporting, and the authors' own limitation statement (Appendix E) undercuts the abstract's meshing claim. These issues need to be resolved before the central claims can be accepted at face value.","major_comments":[{"comment":"The normal estimator Nd is described as 'the maximum eigenvector corresponding to the maximum eigenvalue' of the covariance of KNN points. For a locally planar depth patch, the surface normal is the eigenvector of the smallest eigenvalue; the maximum eigenvector lies in the tangent plane. With τd = 10° (Supp. A.1), the angle θd between a tangent-direction eigenvector and a monocular normal would be near 90° on planar regions, so DNC would mask essentially all depths after step Td = 7k, making the large gains in Table 3 impossible under the stated formulation. If the released code actually uses the minimum eigenvector, the text must be corrected; if it uses the maximum eigenvector, the method cannot work as described. This is load-bearing because DNC is one of the two core contributions.","section":"4.1, Eq. (5)"},{"comment":"Equations (4) and (8) set the filtered depth and normal targets to 0, while the losses in Eqs. (6) and (7) are L1 losses. Under an L1 loss of the form ∥D̂ - 0∥1 or ∥N̂ - 0∥1, masked pixels are pulled toward zero rather than excluded from supervision, which would corrupt rendered depth and normal maps at boundaries and edges—exactly the regions the method claims to filter. The manuscript must specify that a binary mask drops the loss at filtered pixels (or equivalently that the loss is defined only on valid pixels). As written, the mechanism does not implement filtering and is not reproducible.","section":"4.1, Eq. (4) and 4.2, Eq. (8)"},{"comment":"The ScanNet++ evaluation in Table 5 is averaged over only two scenes, and no error bars or significance tests are reported for any of the tables. The abstract claims 'significant improvements' on 'challenging indoor room datasets'; with n = 2 and no variance or per-scene breakdown, the reader cannot judge whether the reported improvements are meaningful. Per-scene results for ScanNet++ and, ideally, error bars or significance tests for MuSHRoom should be provided, or the claim should be tempered to reflect the limited evaluation.","section":"Table 5 and Section 5.1"},{"comment":"Appendix E states: 'The IsoOctree meshing technique we propose focuses on reducing the number of vertices and faces in the mesh while smoothing the surface. However, it does not consistently enhance the overall quality of 3D reconstructions.' This directly qualifies the abstract's claim that the meshing strategy 'recovers finer details from Gaussian models compared to other commonly used open-source meshing tools' and the main-text assertion of 'superior surface creation quality' in Section 5.1. The meshing claim should be narrowed to what the data actually support: the IsoOctree method reduces mesh complexity and can smooth surfaces, but its geometric accuracy is not consistently better than existing TSDF extraction.","section":"Appendix E"},{"comment":"The ANR filter uses normals rendered from the current Gaussian geometry (N̂) to decide which monocular normals (Np) to trust. This is a self-referential loop: early geometry errors can cause the filter to reject correct monocular normals or retain incorrect ones, potentially entrenching the errors. The paper does not provide a diagnostic for this behavior. I would like to see a quantitative report of the fraction of pixels filtered by ANR over training (and, if available, a validation of the filter decisions against the ground-truth normals that ScanNet++ provides), to show that the loop does not simply reinforce the initial Gaussian geometry.","section":"4.2, Eq. (8)"}],"minor_comments":[{"comment":"The word 'inconsitencies' should be 'inconsistencies'.","section":"Section 2"},{"comment":"In the sentence following Eq. (8), the symbol θi should be θn to match the definition in Eq. (5) and the notation θn used earlier.","section":"Section 4.2"},{"comment":"The caption says 'Depth Normal Consistency (DNR)' but the method is abbreviated DNC; this typo should be fixed.","section":"Figure 8 caption"},{"comment":"The text 'We enable our DNR and ANR optimization terms' appears to use 'DNR' where 'DNC' is intended.","section":"Supp. A.1"},{"comment":"The KNN-based normal estimation uses k = 200 world-coordinate neighbors, but no details are given about how the neighborhood is weighted or how noise in the depth map affects the covariance estimate; a brief clarification of the robust estimation would aid reproducibility.","section":"Section 4.1"},{"comment":"The table reports Gaussian counts for a single scene ('vr room'); stating which scene and why this particular count is representative would help the reader interpret the 'reduced Gaussian count' claim.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea—adaptively combining sensor depth and monocular normals for Gaussian Splatting indoor reconstruction—is reasonable and the authors have released code, which is commendable. However, the as-written description of the DNC normal estimation (maximum eigenvector) and the zero-valued filtered targets in the L1 losses are internally inconsistent, and these are not cosmetic issues: they affect the reproducibility of the central contribution. I did not inspect the released code, so it is possible that the implementation uses the minimum eigenvector and a proper mask; if so, the textual errors are fixable. I also note that the ScanNet++ evaluation is very small (two scenes) and the meshing claim is explicitly downplayed by the authors in Appendix E; both should be reconciled with the abstract. Given these load-bearing issues, major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on AGS-Mesh: it's a plausible plug-in for Gaussian splatting that filters noisy phone depth and monocular normals during training, and it shows real gains on indoor scenes. The core idea is that you adaptively trust a prior only when it agrees with the geometry you're optimizing. That's sensible and the experiments mostly back it up. But the write-up has a couple of technical inconsistencies that need fixing before I'd trust the numbers.\n\nWhat's genuinely new: the DNC and ANR filtering losses. DN-Splatter used raw sensor depth and monocular normals; this paper gates those priors by angular consistency. That's a useful extension. The IsoOctree meshing is a nice engineering adaptation of TSDF and octree isosurface extraction, though the gains are modest. The paper is honest in its limitation section that IsoOctree doesn't consistently improve overall mesh quality, which should be reflected in the abstract.\n\nSoft spots, in order of severity:\n\n1. The DNC normal estimation as written is wrong. Section 4.1 says Nd is the maximum eigenvector of the KNN covariance. For a locally planar point cloud, that's the tangent direction, not the normal. With a 10-degree threshold, DNC would mask almost everything after step 7k. The released code probably uses the minimum eigenvector or a different covariance definition, but the paper doesn't say so. This is a reproducibility block.\n\n2. Eqs. (4) and (8) set filtered values to 0. If you plug that into the L1 losses in Eqs. (6) and (7), you'd pull rendered depths and normals toward zero at filtered pixels. There has to be a binary mask in practice, but it's not described. Same reproducibility issue.\n\n3. The evaluation is thin. Only two ScanNet++ scenes, no error bars or significance tests, hand-set thresholds with no sensitivity analysis. The improvement over DN-Splatter on MuSHRoom is real but small (F-score 0.9243 to 0.9433 for Splatfacto+Ours, and 0.9053 to 0.9146 for 2DGS+Ours). That's enough to support the claim that the filtering helps, but not enough to claim \"significant improvements\" loudly.\n\nThe paper deserves a serious referee. The core idea is sound, the ablation shows a clear progression, and the code is supposedly released. But the referee will need the code to resolve the eigenvector and masking questions. If those check out, this is a useful contribution to the 3D reconstruction subfield.\n\nRecommendation: send to peer review. Ask for a revision that fixes the normal estimation description, clarifies the masking, adds error bars on at least the MuSHRoom results, and tones down the meshing claim.","headline":"A sensible adaptive-filtering plug-in for Gaussian splatting that shows real gains, but the DNC normal estimation as written is a reproducibility bug and the meshing claim oversells.","tokens_in":17471,"tokens_out":3027,"would_cite":true,"duration_ms":23710,"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":"Filtering phone depth with monocular normals, and meshing with a depth-aware octree, sharpens indoor 3D reconstructions from Gaussian splatting.","keywords":["3D reconstruction","Gaussian splatting","indoor scenes","depth priors","normal regularization","mesh extraction","octree meshing","smartphone LiDAR"],"falsifier":"Measure the filter's accuracy directly: on a scene with a ground-truth scan, compute the fraction of depth pixels zeroed by Depth Normal Consistency that are actually within a few centimetres of the true surface; if most discarded pixels are correct, the arbitration is wrong. Alternatively, if removing both filters and using raw depth plus oracle normal supervision produces the same F-scores, the claim that consistency-based filtering causes the improvement is falsified.","tokens_in":16321,"feed_emoji":"🏠","tokens_out":8987,"duration_ms":72914,"temperature":0.7,"pith_summary":"Modern indoor 3D reconstruction from smartphones has a tension: phone LiDAR depth is accurate on flat surfaces but poor at edges and far objects, while monocular normal estimates from pretrained networks are sharper at boundaries but inconsistent across views. This paper argues that both signals can be used safely if they are filtered against each other during Gaussian Splatting optimization, rather than trusted wholesale. It introduces two supervision losses, Depth Normal Consistency and Adaptive Normal Regularization, and reports that they improve both mesh metrics and novel-view synthesis on two indoor room datasets captured with smartphones. It also proposes a scale-aware meshing pipeline combining a depth-adaptive truncated signed distance function with octree isosurface extraction to recover finer detail than standard TSDF or Poisson meshing. If these results hold, existing Gaussian Splatting pipelines can be upgraded with plug-in losses and a post-process mesher for more usable room reconstructions from consumer devices.","feed_headline":"Adaptive filtering lifts noisy phone depth to sharper room meshes","feed_subtitle":"Mutual filtering of phone depth and monocular normals improves Gaussian splatting room meshes.","key_machinery":"The argument rests on two adaptive filters and one meshing stage. Depth Normal Consistency computes per-pixel normals from back-projected phone depth by K-nearest-neighbor covariance analysis, compares their angle with monocular normals, and zeroes the depth supervision wherever the angle exceeds a threshold of ten degrees by default, using raw depth early in training and filtered depth later. Adaptive Normal Regularization does the reverse for normals: since monocular normals are inconsistent across views, it first supervises rendered normals with them and, after a number of iterations, retains only monocular normals that agree with normals rendered from the current Gaussian geometry. The meshing machinery is a depth-aware truncated signed distance function, where the truncation distance grows linearly with depth so distant, less reliable observations are weighted less, followed by IsoOctree, an octree-based isosurface extraction that subdivides voxels according to back-projected point density and caps detail where the data do not support it. The isofunction also weights each depth map's contribution by normal-to-ray consistency, and the filters change what the optimization is supervised by, while the mesher changes how the optimized field is converted into a surface.","core_discovery":"On its own terms, the paper establishes that the quality bottleneck in indoor Gaussian Splatting reconstruction is not the absence of geometric priors but the blind use of them. The central discovery is that comparing priors against each other, normals derived from noisy sensor depth versus normals from a monocular network, and monocular normals versus normals rendered from the optimized Gaussians, yields a self-consistent training signal that improves geometry and rendering simultaneously. Concretely, with a ten-degree angle threshold, the Depth Normal Consistency filter removes depth pixels that would drag Gaussian positions off thin structures and edges, while the Adaptive Normal Regularization term suppresses normal supervision where the prior is uncertain, avoiding baked-in multi-view inconsistencies. The paper further claims that geometry extracted from the optimized Gaussian scene is improved by its depth-aware truncated TSDF with IsoOctree extraction, which adapts voxel resolution to point-cloud density and therefore preserves fine detail while smoothing walls. The reported gains are on real smartphone captures from two indoor room datasets, against baselines including volumetric fusion, implicit SDF methods, DN-Splatter, 2DGS, and 3DGS variants, and the method is presented as a plug-in that works for both 3D and 2D Gaussian representations.","pith_inferences":["Editorial inference: the same mutual-consistency filtering could be applied to RGB-only pipelines by using monocular depth as a stand-in for sensor depth, but the paper's own ablation suggests this would be a large quality drop unless the monocular depth is itself calibrated to metric scale.","Editorial inference: the ten-degree thresholds are fixed, and a testable extension would be to make them per-pixel or spatially adaptive, since the paper's own figures show errors concentrate at edges and far surfaces.","Editorial inference: the IsoOctree subdivision heuristic, which expands a voxel when it contains at least fifty back-projected points, ties detail to view coverage; in scenes with highly uneven camera trajectories this might over-smooth rarely seen but geometrically important structures.","Editorial inference: if the Depth Normal Consistency arbitration were wrong on some scenes, the method could still improve on average while discarding correct depth there, so measuring the filter's precision and recall against ground-truth scans would quantify that risk directly."],"forward_implications":["Existing 3D Gaussian Splatting and 2D Gaussian Splatting implementations can add both losses on top of their current training loop, and the paper reports better mesh F-scores and Chamfer distances on room-scale smartphone data without sacrificing rendering quality.","Mesh extraction from a Gaussian scene no longer requires a global voxel resolution: the depth-adaptive TSDF and IsoOctree pipeline recovers fine detail where point density is high and stays smooth elsewhere, at a much smaller mesh size on the reported room scene.","Because the filters depend only on comparing two geometric signals, the method generalizes across 3D and 2D Gaussian variants and is not tied to a particular splatting architecture.","Sensor depth remains load-bearing: the ablation shows raw phone depth supervision greatly outperforms monocular depth supervision, so the method's improvements do not eliminate the need for a physical depth sensor.","Novel-view synthesis from unseen camera trajectories improves with the filtered priors, indicating that the regularization reduces floaters and multi-view inconsistency, not just surface quality."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the first base representation that the plug-in losses are added to.","marker":"[26]"},{"why":"Defines 2D Gaussian Splatting and the rendered-normal-from-depth formula, and is the main framework extended.","marker":"[19]"},{"why":"DN-Splatter is the closest prior method that supervises Gaussians with sensor depth and monocular normals, and serves as the main comparison and base for one implementation.","marker":"[43]"},{"why":"Supplies the pretrained monocular normal estimates used by both filtering strategies.","marker":"[14]"},{"why":"Supplies the octree-based isosurface extraction algorithm used in the proposed meshing stage.","marker":"[24]"},{"why":"Supplies the TSDF and Marching Cubes baselines that the depth-aware TSDF and IsoOctree mesher is compared against.","marker":"[61]"},{"why":"Supplies the smartphone-captured indoor scenes with phone depth and the evaluation protocol used for mesh and novel-view metrics.","marker":"[37]"},{"why":"Supplies the additional large-scale indoor evaluation scenes with high-fidelity reference geometry.","marker":"[55]"}],"fun_headline_variants":["Mutual prior consistency removes noise from phone-based room meshes","Scale-aware meshing recovers fine details from Gaussian room reconstructions","Adaptive depth filtering sharpens Gaussian splatting room meshes","Plug-in depth refinement boosts indoor mesh quality for Gaussian splatting","Adaptive consistency filtering sharpens indoor meshes from noisy phone depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when a normal computed from the phone depth disagrees with a monocular normal estimate by more than ten degrees, the phone depth is the signal to discard, but monocular normals are themselves estimates with their own errors, so a wrong monocular normal would make the filter throw away correct depth.","fun_headline_variants_meta":{"raw":{"variants":["Mutual prior consistency removes noise from phone-based room meshes","Scale-aware meshing recovers fine details from Gaussian room reconstructions","Adaptive depth filtering sharpens Gaussian splatting room meshes","Plug-in depth refinement boosts indoor mesh quality for Gaussian splatting","Adaptive consistency filtering sharpens indoor meshes from noisy phone depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001241,"raw_usage":{"total_tokens":5142,"prompt_tokens":1044,"completion_tokens":4098,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":4008}},"tokens_in":660,"tokens_out":4098,"duration_ms":24331,"temperature":1.0,"reasoning_tokens":4008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:20:27.575669+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the filter's accuracy directly: on a scene with a ground-truth scan, compute the fraction of depth pixels zeroed by Depth Normal Consistency that are actually within a few centimetres of the true surface; if most discarded pixels are correct, the arbitration is wrong. Alternatively, if removing both filters and using raw depth plus oracle normal supervision produces the same F-scores, the claim that consistency-based filtering causes the improvement is falsified.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines 3D Gaussian Splatting, the first base representation that the plug-in losses are added to."},{"cited_title":"2d gaussian splatting for geometrically accu- rate radiance fields","cited_arxiv_id":null,"evidence_quote":"Defines 2D Gaussian Splatting and the rendered-normal-from-depth formula, and is the main framework extended."},{"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 pretrained monocular normal estimates used by both filtering strategies."},{"cited_title":"Unconstrained isosurface extraction on arbitrary oc- trees","cited_arxiv_id":null,"evidence_quote":"Supplies the octree-based isosurface extraction algorithm used in the proposed meshing stage."},{"cited_title":"Mushroom: Multi-sensor hybrid room dataset for joint 3d reconstruction and novel view synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the smartphone-captured indoor scenes with phone depth and the evaluation protocol used for mesh and novel-view metrics."},{"cited_title":"Scannet++: A high-fidelity dataset of 3d indoor scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the additional large-scale indoor evaluation scenes with high-fidelity reference geometry."}],"review_version":1}