{"id":"8d85156f-08e6-4774-80ca-442362436a40","arxiv_id":"2412.03910","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid of deformable Gaussian splatting and dynamic neural SDF achieves state-of-the-art 3D mesh accuracy from monocular video while keeping view synthesis competitive.","lead":"This paper presents DGNS, a method that reconstructs both the 3D shape and the novel views of a moving scene from a single monocular video by combining deformable 3D Gaussian splatting with a dynamic neural surface. It reports the best 3D reconstruction accuracy on two public dynamic-scene benchmarks while matching the top rendering quality, though the published depth-filtering equation appears to contain a sign error and no code is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) as printed computes (d_alpha - d_m)/2, which is near zero when the two depths agree; Eq. (6) then supervises the SDF near the camera, so the paper's central depth-filtering contribution is not reproducible as written.","rationale":"The reader identified the same load-bearing concern: Eq. (5) as printed is internally inconsistent with the surrounding text and figures. My stress-test confirms this is the most direct threat to the paper's central claim. The depth filter is a named contribution and the principal source of SDF supervision; if the formula is wrong as written, the method is not reproducible, and the ablation evidence for the benefit of filtered depth is called into question. I considered the homeomorphic-deformation-versus-topological-change contradiction raised in the reader's rationale. It is a genuine theoretical tension, but it is less decisive here because the final meshes are extracted from the DGS module via GauStudio, so topological changes may be handled outside the homeomorphic DNS mapping. The Eq. (5) issue is more load-bearing because it affects the core supervision mechanism directly. The likely explanation is a sign typo rather than a substantive methodological flaw, so the appropriate disposition remains CONDITIONAL: the authors should correct the equation, release code or a precise derivation, and re-confirm the ablation. I agree with the reader's verdict and would not change it.","tokens_in":16469,"tokens_out":3549,"duration_ms":34913,"concrete_test":"Re-implement the DNS depth-supervision branch exactly as published: set d_f = (d_alpha - d_m)/2 for |d_alpha - d_m| < tau_f and use Eq. (6) on one Dg-mesh sequence (e.g., Torus2sphere). If the filtered-depth ablation (Table 3 row 'Filtered Depth') cannot reproduce CD=0.697 / EMD=0.121 / PSNR=35.02 and instead collapses (because d_f is near zero), then the published Eq. (5) is not the operation that produced the results. The paper must be corrected to the midpoint formula (d_alpha + d_m)/2 and the ablation re-validated before the depth-supervision claim can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The depth-filtering rule is load-bearing because it feeds the SDF supervision Lsdf in Eq. (6), and Table 3 attributes a large part of the reported performance to filtered depth. As printed, Eq. (5) returns d_f = (d_alpha - d_m)/2 whenever |d_alpha - d_m| < tau_f. Since the filtering condition forces the two depth estimates to be nearly equal, d_f is approximately 0 for every accepted ray. Substituting into Eq. (6) gives Lsdf = sum ||F(H(o + d_f * v, t))||_1, which supervises the SDF at points essentially at the camera origin, not on the surface. This contradicts Fig. 3, which shows filtered points on the object surface, and contradicts the ablation gains in Table 3. The intended operation is very likely the midpoint (d_alpha + d_m)/2, and the printed minus sign is a typo. But the paper as published contains no code or supplementary derivation to resolve the ambiguity; a reader implementing Eq. (5) literally cannot reproduce the method. Because the depth-supervision mechanism is one of the three named interactions between the DGS and DNS modules, the central claim of state-of-the-art geometry with competitive rendering rests on an equation that cannot be correct as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGNS, a hybrid framework that jointly optimizes a deformable 3D Gaussian splatting (DGS) module for view synthesis and a dynamic neural surface (DNS) module for geometry reconstruction in monocular dynamic scenes. The two modules interact in three ways: DGS-rendered depth maps guide DNS ray sampling and provide depth supervision, DNS SDF values guide Gaussian growth/pruning via surface-aware density control, and a foundation-model normal prior regularizes both modules. Experiments on the Dg-mesh, D-NeRF, and Nerfies datasets report state-of-the-art Chamfer Distance and EMD on Dg-mesh while remaining competitive with D3DGS in novel-view synthesis.","tokens_in":16787,"tokens_out":6091,"duration_ms":52971,"significance":"If the reported results hold, the paper makes a useful contribution by demonstrating a concrete mechanism for coupling an explicit Gaussian renderer with an implicit dynamic SDF, and by showing that geometry-aware depth supervision can improve dynamic reconstruction without sacrificing rendering quality. The evaluation against external ground-truth meshes on Dg-mesh is a strength, and the ablation study in Table 3 gives initial evidence for the contribution of filtered depth and normal supervision. However, the central depth-filtering equation is internally inconsistent as printed, several hyperparameters are unreported, and all quantitative claims rest on single-run point estimates, so the current manuscript does not yet substantiate the headline state-of-the-art claim in a reproducible way.","major_comments":[{"comment":"The depth-filtering rule as printed is internally inconsistent with its stated purpose. In the accepted branch, d_f = (d_alpha - d_m)/2, and the acceptance condition |d_alpha - d_m| < tau_f forces this quantity to be near zero for every accepted ray. Substituting this d_f into Eq. (6) yields L_sdf = sum ||F(H(o + d_f v, t))||_1 evaluated essentially at the camera origin, not on the object surface. This contradicts Fig. 3, where filtered points lie on the reconstructed surface, and it contradicts the ablation in Table 3, which attributes a large CD improvement (0.746 to 0.502) to filtered depth. If the intended operation is the midpoint (d_alpha + d_m)/2, the printed minus sign is a typo; however, no code or supplementary derivation is provided, so the central depth-supervision mechanism is not reproducible from the text as written. Please correct Eq. (5), report the value of tau_f, and clarify the behavior of the loss for rejected rays.","section":"§4.1, Eq. (5)"},{"comment":"All quantitative results are point estimates without error bars, seeds, or a statement of the number of runs. The headline claim in §5.2 that the method is 'unique in offering consistently superior performance' rests on differences that are sometimes very small (e.g., Duck CD 0.773 vs. 0.790 and EMD 0.046 vs. 0.047; several D-NeRF PSNR values are below D3DGS, e.g., Mutant 41.47 vs. 42.63 and Hook 36.34 vs. 37.42). Without multiple runs or a variance estimate, the claimed state-of-the-art status is not established at the reported precision. Reporting mean and standard deviation over at least three seeds, or otherwise justifying that the observed differences exceed run-to-run noise, is needed for the central claim.","section":"§5.1 and Tables 1-3"},{"comment":"Several hyperparameters that determine the method are not reported: tau_f in Eq. (5), w_g, w_p, tau_g, and tau_p in Eqs. (8)-(9), and lambda_sdf, lambda_nn, and lambda_eik in Eq. (12). Only s and tau_d are given in §5.1. These parameters control the surface-aware density control and the SDF/normal/eikonal losses, which are central to the hybrid interaction. The manuscript should provide their values or a clear pointer to released code so that the experiments can be reproduced.","section":"§5.1 and Eqs. (8), (9), (12)"}],"minor_comments":[{"comment":"The training schedule is stated twice with conflicting numbers: 'warm-up phase (0 to 10k iterations) followed by joint training (10k to 40k iterations)' and later 'warm-up phase (0 to 15k iterations) followed by joint training (15k to 40k iterations)'. Please reconcile these two descriptions.","section":"§5.1, Implementations"},{"comment":"The text says 'Marigold [24, 38]', but [24] is a monocular depth-estimation paper and [38] is a diffusion-fine-tuning paper; the source of the pseudo-normal maps is unclear. Please cite the correct normal-prediction model or clarify how the depth model is used to produce normals.","section":"§4.1, Normal supervision"},{"comment":"The notation d(x) is introduced in the sentence after Eq. (2) but is not used in the equations; please align the notation so that the SDF is consistently denoted.","section":"§3.2, Eqs. (2)-(3)"},{"comment":"The alpha-depth formula should be written more conventionally; as printed, the denominator is the sum of weights, but the expression can be simplified and should be checked to confirm that it corresponds to the intended weighted average depth.","section":"§4.1, Eq. (4)"},{"comment":"The caption says 'images from left to right are 3D point clouds projected from alpha-blending depth, median depth, and filtered depth' but does not explicitly identify subfigures (a)-(d) or the RGB image; please clarify the correspondence.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case. The proposed idea is coherent and the reported numbers are plausible, but the published text lacks the information needed to verify the central depth-filtering step, and the experimental evidence is all single-run. I would ask for a corrected Eq. (5), a complete hyperparameter table, and at least a small set of repeated runs on the most contested metrics before acceptance. The paper is within the scope of ACM MM, but the absence of released code makes the ambiguities more consequential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The hybrid architecture is the real contribution: deformable Gaussians for appearance, a dynamic neural SDF for geometry, with depth from the Gaussians guiding SDF ray sampling and supervision, and the SDF pulling Gaussian density control toward the surface. That coupling is genuinely new in this form, and the Dg-mesh results support it: DGNS gets the lowest Chamfer distance and EMD on most objects while staying within a few tenths of a dB of D3DGS in PSNR. If those numbers hold, this is a useful baseline for anyone who needs both mesh quality and view synthesis from monocular video.\n\nNow the soft spots, in order of seriousness. First, Eq. (5) as printed is not a depth filter: it computes (d_alpha - d_m)/2, and because the condition requires the two depths to be nearly equal, the output is always near zero. That would supervise the SDF at the camera, not the surface. Fig. 3 and the Table 3 ablation make it obvious the authors meant (d_alpha + d_m)/2. I think this is a sign typo, not a conceptual failure, but it is load-bearing: the depth-supervision mechanism is one of the three named DGS-to-DNS interactions, and without the corrected equation the paper is not reproducible. The stress-test note lands.\n\nSecond, the evaluation is single-run with no error bars, and several important hyperparameters (tau_f, tau_d, w_g, w_p, tau_g, tau_p, and the loss weights) are never given concrete values in the text. The implementation paragraph mentions s and lambda_I, but not the rest. That matters because the ablations show large swings from depth and normal terms, and the reader has no way to know how sensitive the results are.\n\nThird, the paper claims topological robustness in Fig. 5, but the SDF deformation field is described as homeomorphic, which by definition cannot handle topological change. That is a minor inconsistency in the text unless the authors clarify how the meshes actually handle topology.\n\nThe mutual bootstrapping concern is real but not damning. Depth from DGS supervises the SDF, and the SDF guides Gaussian placement; that is a cooperative optimization rather than a circular argument, but it does mean the two modules are not independently validated. The ground-truth mesh evaluation on Dg-mesh is external, so the headline geometry claim is not fitted to the metric, which is good.\n\nBottom line: this is a competent engineering paper with a plausible idea and strong empirical results. The typo in Eq. (5) must be fixed, hyperparameters and error bars added, and ideally code released. I would send it to peer review rather than desk reject, because the architecture is worth examining and the flaws are fixable. A careful referee should focus on the corrected depth filter and on whether the reported gains survive a reimplementation.","headline":"Useful hybrid for dynamic reconstruction, but the printed depth-filter equation is wrong as written; fix it and this is a solid baseline.","tokens_in":17325,"tokens_out":2142,"would_cite":false,"duration_ms":20536,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-module hybrid of deformable Gaussians and a dynamic neural surface claims state-of-the-art dynamic 3D geometry with rendering quality intact.","keywords":["dynamic scene reconstruction","3D Gaussian Splatting","neural signed distance function","novel view synthesis","monocular video","depth supervision","surface-aware density control","deformable neural surface"],"falsifier":"Run the published Eq. (5) exactly on a ray where $|d_\\alpha-d_m|<\\tau_f$; the filtered depth collapses to near zero, and Eq. (6) then penalizes the SDF at the camera origin, which would destroy the reconstruction. Inspecting the released code to see whether it computes $(d_\\alpha+d_m)/2$ instead settles whether the claimed depth-supervision mechanism exists.","tokens_in":16268,"feed_emoji":"🎥","tokens_out":7997,"duration_ms":154158,"temperature":0.7,"pith_summary":"DGNS claims that monocular dynamic 3D reconstruction does not have to choose between accurate geometry and high-quality rendering. It couples a deformable Gaussian splatting module, which handles appearance, with a dynamic neural signed-distance surface, which handles shape. The Gaussian module's rendered depths guide the surface module's ray sampling and supervise its SDF, while the surface's zero-level set steers where Gaussians are split, cloned, and pruned. On the Dg-mesh benchmark the paper reports the lowest Chamfer Distance and Earth Mover's Distance on most objects while staying competitive in PSNR, and on D-NeRF it keeps rendering quality near the best baseline while producing smoother meshes. If true, the practical payoff is a single video-to-mesh-and-view pipeline for articulate and deforming scenes.","feed_headline":"Gaussian-SDF hybrid nails dynamic 3D geometry, keeps view quality","feed_subtitle":"Depth from deformable Gaussians supervises a neural surface, which in turn anchors the Gaussians to that surface.","key_machinery":"The load-bearing machinery is bidirectional depth-and-surface coupling between two modules that share a deformation field. The DGS module renders an alpha-blended depth and a median depth; a filter keeps a depth only when the two agree, and this filtered depth both centers the SDF ray-sampling interval and supplies the SDF regression target of Eq. (6). Conversely, the DNS zero-level set is converted into a Gaussian falloff signal that is added to the gradient criterion for splitting and cloning and subtracts from opacity for pruning, so Gaussian density follows the reconstructed surface. A monocular normal prior regularizes both modules. The named device \"surface-aware density control\" is what lets the explicit Gaussians inherit geometry without sacrificing the appearance fidelity of pure splatting.","core_discovery":"The central claim is that a two-module hybrid named DGNS beats both pure implicit surface methods and pure deformable-Gaussian methods because each representation supplies what the other lacks. Deformable 3D Gaussian splatting gives dense but noisy depth maps near the true surface, and those depths concentrate neural SDF ray sampling and provide supervision that pulls the zero-level set into place. The neural SDF in turn returns a geometry signal used to grow and prune Gaussians so they sit on the surface instead of floating in space. On Dg-mesh the paper reports the best reconstruction errors on nearly every object (for example, Chamfer Distance 0.773 versus 0.790 for the closest baseline on Duck, 0.289 versus 0.299 on Horse, and 0.413 versus 0.482 on Girlwalk) while PSNR stays competitive, and on D-NeRF DGNS is within a few tenths of a decibel of the rendering leader while producing qualitatively smoother meshes.","pith_inferences":["The printed Eq. (5) appears to contain a sign typo: under its own condition $|d_\\alpha-d_m|<\\tau_f$, the formula $(d_\\alpha-d_m)/2$ makes the filtered depth nearly zero, so Eq. (6) would supervise the SDF at the camera origin rather than at the surface; the intended filter is most likely the midpoint $(d_\\alpha+d_m)/2$.","The same alpha-plus-median depth filter could be lifted out as a general denoising step for any Gaussian-splatting depth map before it is used as supervision, independent of the neural surface module.","Surface-aware density control is a transferable recipe: any Gaussian-SDF hybrid, static or dynamic, could use the SDF zero-level set to decide where to add and remove primitives.","The quantitative geometry evidence is strongest where ground-truth meshes exist, namely Dg-mesh; on the real Nerfies sequence the evidence is qualitative, so behavior under real-world noise in monocular depths and normals is not yet quantified by the paper."],"forward_implications":["A single monocular video can yield both a frame-consistent deforming mesh and photorealistic novel views, closing the geometry-versus-rendering tradeoff that separates implicit-only and Gaussian-only methods.","Depth rendered by deformable Gaussians can serve as a cheap guiding signal for neural SDF training: it shortens ray marching and anchors SDF supervision, so the surface module needs less blind search to converge.","The SDF's zero-level set can be used as a principled prior for Gaussian split, clone, and prune decisions, reducing the floaters that plague splatting-based surface reconstruction.","Monocular normal priors from a pretrained foundation model improve both appearance and geometry in under-constrained dynamic scenes, with the ablation reporting Chamfer Distance 1.006 with neither cue versus 0.502 with filtered depth and normals together.","On both the Dg-mesh and D-NeRF benchmarks, the hybrid's rendering PSNR remains competitive with the strongest deformable-Gaussian baseline, so the geometry gains do not come at a perceptual cost."],"supporting_citations":[{"why":"Supplies the 3D Gaussian splatting formulation, including primitives, alpha-blending rasterization, and adaptive density control, that the DGS module extends.","marker":"[25]"},{"why":"Introduces the deformable Gaussian deformation field used by DGS and serves as the strongest rendering baseline DGNS must match.","marker":"[63]"},{"why":"Provides the Dg-mesh monocular dynamic dataset, the Gaussian-mesh anchoring baseline, and the ground-truth meshes used for CD and EMD evaluation.","marker":"[31]"},{"why":"Gives the invertible deformation-field plus neural SDF formulation that the DNS module is built on and is also a geometry-focused baseline.","marker":"[3]"},{"why":"Supplies the static-scene surface-aware density control strategy that DGNS adapts so SDF guides Gaussian growth and pruning.","marker":"[68]"},{"why":"Provides the D-NeRF dynamic-scene benchmark and one of the main baselines for novel-view synthesis comparison.","marker":"[45]"},{"why":"Supplies the pretrained monocular normal foundation model whose pseudo-normals regularize both the GS and SDF modules.","marker":"[24]"},{"why":"Defines the volume-rendered normal consistency loss used to supervise the neural surface's normals.","marker":"[49]"}],"fun_headline_variants":["Hybrid Gaussians + neural surface: best dynamic 3D meshes yet","DGNS fuses Gaussians and SDF for sharper dynamic geometry","Dual representation: Gaussians guide SDF, SDF anchors Gaussians","Monocular dynamic 3D: best mesh via Gaussian-depth-guided SDF","Depth-supervised SDF from Gaussians tops dynamic 3D reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the depth-filtering rule in Eq. (5) actually identifies the surface: as printed, when the alpha-blended and median depths are close the rule outputs about zero, so Eq. (6) would supervise the SDF at the camera rather than at the surface. If the intended midpoint average is not what the implementation uses, the depth-supervision contribution is not reproducible from the paper.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid Gaussians + neural surface: best dynamic 3D meshes yet","DGNS fuses Gaussians and SDF for sharper dynamic geometry","Dual representation: Gaussians guide SDF, SDF anchors Gaussians","Monocular dynamic 3D: best mesh via Gaussian-depth-guided SDF","Depth-supervised SDF from Gaussians tops dynamic 3D reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00112,"raw_usage":{"total_tokens":4637,"prompt_tokens":901,"completion_tokens":3736,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":3637}},"tokens_in":517,"tokens_out":3736,"duration_ms":23007,"temperature":1.0,"reasoning_tokens":3637,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:57:40.875900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published Eq. (5) exactly on a ray where $|d_\\alpha-d_m|<\\tau_f$; the filtered depth collapses to near zero, and Eq. (6) then penalizes the SDF at the camera origin, which would destroy the reconstruction. Inspecting the released code to see whether it computes $(d_\\alpha+d_m)/2$ instead settles whether the claimed depth-supervision mechanism exists.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the invertible deformation-field plus neural SDF formulation that the DNS module is built on and is also a geometry-focused baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained monocular normal foundation model whose pseudo-normals regularize both the GS and SDF modules."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the volume-rendered normal consistency loss used to supervise the neural surface's normals."}],"review_version":1}