{"id":"4a07ba87-afe9-4c1f-a696-dd8838abd414","arxiv_id":"2501.08577","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A scene is reconstructed as a graph of overlapping local neural SDFs that are registered and blended, improving detail and enabling large-scale reconstruction.","lead":"The authors split a 3D scene into overlapping pieces, reconstruct each piece with its own neural signed distance function, then align and blend the pieces into one model. The method aims to produce more detailed surfaces and to scale to large areas that a single neural network cannot handle.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Error accumulation along the MST propagation chain is unquantified, and the registration optimization is self-supervised, so the city-scale scalability claim rests on an unvalidated assumption.","rationale":"The reader's weakest_assumption identified exactly the registration assumption as load-bearing: the global similarity transform model and the ability of the optimization to compensate for residuals. My attack agrees and sharpens it. The paper's own numbers—24 edges in the MST, only 5 edges reported in Table 6—make error accumulation the concrete risk. The strongest claim of scalable high-quality reconstruction depends on the MST chain accumulating only negligible error, and the paper provides no direct test of that condition. The self-supervised nature of the color-based refinement makes the absence of external geometric validation particularly salient. I agree with the CONDITIONAL verdict: the core idea is coherent, and the high-quality claims on small objects are supported by quantitative metrics against ground truth. However, the city-scale scalability claim requires additional evidence. The proposed loop-closure and keypoint-contrast tests directly measure error accumulation without needing ground-truth mesh, and would settle whether the concern actually lands. This is a good-faith critique focused on missing evidence rather than an internal inconsistency.","tokens_in":22859,"tokens_out":1549,"duration_ms":15704,"concrete_test":"Run a known-sequence experiment on the Campus dataset: after computing all 24 edge transformations and the global alignment, independently verify alignment using loop constraints—for each of, say, 5 cycles in the overlap graph, compose the transforms around the cycle and measure the deviation from identity in translation, rotation, and scale. If the maximum cycle error exceeds a threshold (e.g., more than 0.5% of the scene bounding-box diagonal, or rotation > 0.1 degrees), the MST propagation is demonstrably drifting and the city-scale reconstruction claim needs re-evaluation. Additionally, compare a small set of shared image keypoints' 3D positions after global alignment against the per-node COLMAP point clouds to quantify residual misalignment independent of the color-based loss.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central scalability claim depends on the MST propagation chain: the global coordinate system for node 25 is obtained by composing 24 pairwise similarity transforms. Section 3.3 analyzes only a single edge, and Table 6 reports rendering metrics for only 5 of the 24 Campus edges. If each edge retains a small residual error, errors compound along the chain; the paper provides no analysis of how rotation, translation, and scale residuals propagate through the composition T_global = T_0k ∘ ... ∘ T_ij. Critically, the registration optimization minimizes the color loss ||C_j(P_i · T_ij) − C_j(P_j)||, which is self-supervised using rendered colors of the target node rather than ground-truth images or geometric correspondences. A local minimum can satisfy this loss while leaving a systematic geometric bias, such as a small rotation or scale error, and the paper does not validate alignment against any independent geometric ground truth (e.g., fixed distances between buildings, a global COLMAP alignment, or loop-closure constraints). The large-scale qualitative results in Figures 10–11 are visually impressive but cannot reveal small misalignments, and the MST chain can accumulate them. Additionally, the initial registration in Eq. (8) assumes a single similarity transform relating the two node world coordinate systems; if per-node COLMAP reconstructions contain non-rigid drift or varying scale within the overlap region, this initial guess is biased and the color-based refinement cannot correct a globally wrong scale. No error bars, no reported failure cases, and no code release further weaken the scalability evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural SDF-Graph, a divide-and-conquer representation for neural implicit surface reconstruction. A scene is decomposed into overlapping local SDF nodes, each reconstructed independently with an existing method such as NeuS or Bakedangelo; adjacent nodes are registered using a least-squares similarity transform initialized from shared COLMAP poses and then refined by a volume-rendering color loss; finally the local SDFs are combined with a softmax-weighted blending scheme. Experiments on Lego, Jade, BlendedMVS, Actors-HQ, and two drone-captured campus datasets report improved Chamfer distance and F-score relative to single-network baselines, and demonstrate scaling to 25 nodes over a roughly 1200m x 800m area, along with texture generation and scene editing applications.","tokens_in":23157,"tokens_out":3593,"duration_ms":33583,"significance":"The central claim—that partitioning a scene into multiple overlapping local SDFs improves reconstruction fidelity and enables scalable reconstruction—is supported by consistent quantitative gains on external benchmarks with ground-truth meshes, e.g., Lego Chamfer distance improving from 5.209e-4 (baseline) to 2.836e-4 (8 nodes) in Table 1 and Jade from 2.507e-4 to 0.958e-4 (4 nodes) in Table 2. The paper also includes a parameter-count ablation (Table 7) and a comparison against recent 3DGS-based reconstructions, and it uses publicly available components (COLMAP, NeuS, SDFStudio), which helps reproducibility. The main open risk is the validation of the registration chain underlying the city-scale claim: only five of the twenty-four Campus edges are reported in Table 6, the registration optimization is self-supervised with no independent geometric ground truth, and the effect of error accumulation along the MST is not quantified.","major_comments":[{"comment":"The city-scale claim rests on composing 24 pairwise similarity transforms along the MST to obtain a global coordinate system, but the paper reports rendering metrics for only 5 of the 24 edges (Table 6) and provides no analysis of how rotation, translation, and scale residuals propagate through the composition. Please add either a global geometric validation (e.g., distances between fixed landmarks, comparison with a global COLMAP alignment, or loop-closure errors) or a quantitative residual analysis along the full chain; without this, the absence of visible seams in Figures 10–11 cannot rule out accumulated misalignment.","section":"Sec. 3.3 and Sec. 4.4"},{"comment":"The registration optimization minimizes the color difference between two renderings from the target node's color field C_j, and the Table 6 metrics are computed against those same rendered images. This is a self-supervised validation: a local minimum of Eq. (12) can leave a systematic geometric bias (e.g., a small rotation or scale error) that rendering metrics would not reveal. Please validate the final alignment with an independent geometric criterion on at least one scene, such as a global bundle adjustment or manually measured control points.","section":"Sec. 3.3, Eq. (12)"},{"comment":"The initial registration in Eq. (8) assumes that the two per-node COLMAP reconstructions are related by a single similarity transform over the whole overlap region. If the per-node reconstructions contain non-rigid drift or spatially varying scale, the least-squares initial guess is biased and the color-based refinement may not compensate. The manuscript does not analyze the magnitude of such residuals or the sensitivity of the final reconstruction to them; please add an experiment that perturbs the initial transform and measures the resulting registration error, or report drift statistics over the overlap regions.","section":"Sec. 3.3, Eqs. (4)-(9)"},{"comment":"All quantitative claims, including the headline '45.6% improvement in Lego Chamfer distance,' are based on single runs with no error bars or multiple random seeds. Given that the core assertion is a consistent quality gain from increasing the number of nodes, please provide at least three runs per configuration or otherwise report the variance, so that the significance of the gains relative to training stochasticity can be assessed.","section":"Sec. 4.2, Tables 1-3"}],"minor_comments":[{"comment":"There are several typos and grammatical slips, e.g., 'loacl' (Sec. 1), 'ddemonstrate' (Sec. 4.3), 'basline' (Sec. 4.3), 'specfied' (Sec. 3.4), and 'T raining' (Appendix A.1.1).","section":"Throughout"},{"comment":"The summation index and matrix dimension use 'N12' in Eq. (8), but the text defines the number of shared images as Nij; please correct the notation.","section":"Sec. 3.3, Eq. (8)"},{"comment":"The reported overlap percentages for the Lego two-node division (41.67%, 24.75%, 18.38%) are stated without explaining which denominator each percentage refers to; please clarify the definitions of 'lower node range,' 'upper node range,' and 'global range.'","section":"Sec. 4.2"},{"comment":"The COLMAP runtimes for the four Sub-Campus nodes sum to 5h2m (1h47m + 58m + 1h33m + 44m), but the table reports 'Total: 5h8m'; please check the arithmetic.","section":"Table 5"},{"comment":"Figure 14 shows ablation results labeled 'Lego,' 'Campus,' and 'Human body,' but the text of Section 4.6 appears to describe only the Lego case; please clarify which dataset each quantitative pair corresponds to.","section":"Fig. 14 and Sec. 4.6"},{"comment":"The statement that each overlapping node pair shares 'around 100' images is vague; please provide a range or a small table of shared-image counts for the Campus nodes.","section":"Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and the core divide-and-conquer idea is likely to interest the community, but the city-scale scalability claim currently depends on registration validation that is largely self-referential and incomplete. I recommend major revision with a request for independent geometric validation of the registration chain and statistical robustness of the headline quantitative comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper applies a known divide-and-conquer strategy from NeRF to neural SDF geometry, and the central claim that local SDFs beat a single global SDF is supported by external ground-truth metrics. The scalability claim, however, rests on evidence that is partly self-supervised and unquantified. It deserves a serious referee, but with a clear request for more rigorous validation.\n\nWhat is actually new: Block-NeRF and Mega-NeRF did block decomposition for color radiance fields; this paper does it for signed distance geometry, with an SDF graph, MST-based registration, and a softmax SDF blending that handles seams. The blending trick is simple and looks effective. The quantitative results on Lego and Jade are convincing: eight nodes cut Chamfer distance by roughly 45% on Lego, and F-scores improve monotonically. The parameter-count ablation is a nice touch—it shows quadrupling a single network's capacity does not get you what divide-and-conquer provides. Consistent gains on BlendedMVS and the human body, all against ground-truth meshes, make the core claim credible.\n\nWhere it gets soft: the registration validation is circular in a specific sense. The optimization in Eq. (12) minimizes color differences against the target node's own rendered colors, and Table 6 reports that the 'final' metrics approach the 'target' metrics. That mostly tells you the optimization minimized its own objective; it does not independently verify geometric alignment. A color-consistent local minimum can hide a systematic rotation or scale error. The paper gives no independent geometric ground truth for the campus registration, no analysis of error accumulation along the 24-edge MST chain, and no failure cases. The large-scale reconstruction is also only qualitative—no ground truth for a 1200m by 800m campus, so visual plausibility carries a lot of weight. Add the lack of error bars and code release, and the scalability claim is genuinely under-supported, even if plausible.\n\nNone of this sinks the main reconstruction-quality claim, which stands on external benchmarks. But the paper oversells 'urban-scale' scalability relative to the evidence. A good revision would add independent validation of registration (e.g., fixed distances between buildings, comparison with global COLMAP on a smaller scene, or per-edge residual statistics) and quantify MST error accumulation on a scene with ground truth.\n\nThis is a practical contribution for people working on large-scale neural implicit reconstruction, especially in mapping and urban modeling. Yes, send it to peer review—it is substantive enough to deserve referee time, but reviewers should push hard on the missing registration validation.","headline":"A solid engineering contribution: the core claim that local SDFs beat a global one survives scrutiny, but the scalability story is under-validated and the registration evaluation is partly self-referential.","tokens_in":23677,"tokens_out":2963,"would_cite":true,"duration_ms":30408,"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":"Modeling a scene as overlapping local neural SDFs, registered and blended, yields higher-fidelity and scalable 3D reconstruction than a single global network.","keywords":["signed distance function","neural implicit surface","3D reconstruction","scene graph","registration","SDF blending","scalable reconstruction","divide-and-conquer"],"falsifier":"Take a pair of overlapping nodes from a large scene, apply a smooth non-rigid warp (for example, a 5% scale gradient across the overlap) to one node's structure-from-motion output, then run the full registration; if the optimized color-consistency loss converges to a visible seam or a wrong transform, the single-similarity-transform assumption is falsified.","tokens_in":22655,"feed_emoji":"🧩","tokens_out":11095,"duration_ms":102243,"temperature":0.7,"pith_summary":"The paper argues that the standard neural-SDF approach—one global multi-layer perceptron representing the whole scene—limits both accuracy and scale, and it proposes a divide-and-conquer representation instead: a graph of independent local SDFs with overlapping regions that are reconstructed separately, registered using shared images, and blended into one global SDF. The intended payoff is that each local network concentrates its capacity on a small region, so fine detail improves while the scene can be extended by adding nodes. On the Lego scene the Chamfer distance improves by 45.6% going from the single-network baseline to eight nodes, and the same pipeline reconstructs a 25-node campus scene covering roughly 1200 m by 800 m. If the representation works as claimed, it is a practical route to urban-scale neural surface reconstruction without a single global pose optimization.","feed_headline":"Eight local neural SDFs cut Lego reconstruction error by 45.6%","feed_subtitle":"A graph of overlapping local SDFs also scales to a 1200×800 m campus reconstruction.","key_machinery":"The signed distance function (SDF) is the function whose zero set is the surface, and each graph node stores one local SDF. Two mechanisms carry the argument. First, overlap-based registration: because shared images see the same 3D point in both local coordinate systems, the relation $P^i_h \\mathcal{T}'_{ij}=P^j_h$ yields an overdetermined linear system whose least-squares solution is an initial similarity transform; this transform is then optimized by volume-rendering the registered node's color field in the reference node and minimizing the color difference over the shared-image masks. Second, softmax SDF blending: instead of taking $\\min(f_i(x),f_j(x))$ in overlapping regions, the method averages with weights that decay with distance to each node's boundary, eliminating the seam while staying close to a valid SDF. The graph's minimum spanning tree is used so that each pair of adjacent nodes is registered only once and the global coordinate frame is propagated from node to node.","core_discovery":"The central claim is that a single global MLP is the bottleneck in neural SDF reconstruction, and that representing a scene as a Neural SDF-Graph removes that bottleneck. The graph $G=(V,E)$ has one node $v_k$ per local SDF, each reconstructed from its own subset of images and its own camera poses; an edge $e_{ij}$ marks two nodes that share images and therefore overlap in space. Adjacent nodes are aligned by an initial similarity transform $\\mathcal{T}'_{ij}$ (rotation, translation, scale) obtained in closed form from the shared camera poses, then refined by minimizing a color-consistency loss through volume rendering. The aligned local SDFs are fused with softmax weights $\\omega^k_{ij}(x)=e^{-\\beta d_k(x)}/(e^{-\\beta d_i(x)}+e^{-\\beta d_j(x)})$, where $d_k(x)$ is the distance to the boundary of node $v_k$, producing a global SDF without the seam that a plain min-union creates. The quantitative evidence is that reconstruction error falls steadily as node count rises—45.6% better Chamfer distance on Lego with eight nodes than the single-network baseline—and that the same recipe handles a 25-node, roughly 1200 m by 800 m campus.","pith_inferences":["Not stated in the paper: the registration step is agnostic to what the local field represents, so the same overlapping-node graph could wrap explicit representations such as 3D Gaussian splatting or mesh blocks; the softmax blending would need an SDF-like overlap analogue.","Not stated in the paper: because the global frame is propagated along a minimum spanning tree, any residual registration error on an early edge is inherited by all later nodes, so a loop-closure or global bundle-adjustment pass over the whole graph is the natural stress test for very large scenes.","Not stated in the paper: the overlap fraction is chosen by hand in the experiments (roughly 15-42% for Lego divisions, 20% for Jade), so an automatic schedule that trades registration reliability against wasted training is a testable extension.","Not stated in the paper: the authors note that neural SDF methods assume watertight surfaces; a non-watertight local SDF model would remove the irregular iso-surface shapes they observe near node boundaries."],"forward_implications":["With more local nodes, reconstruction detail improves monotonically in the reported tests: the Lego Chamfer distance drops from $5.209\\times 10^{-4}$ for a single network to $2.836\\times 10^{-4}$ with eight nodes.","Scene size is no longer bounded by one network: adding nodes along the data-collection path produced a 25-node reconstruction of a roughly 1200 m by 800 m campus from 5973 oblique aerial images.","Per-node camera estimation is much cheaper than global pose estimation: on the Sub-Campus scene, structure-from-motion time fell from 14 h 25 min to 5 h 8 min.","The independence of nodes supports downstream uses: each node's mesh and texture can be extracted and optimized separately, and a node's registration transformation can be changed to move objects within the scene.","Softmax-based SDF blending removes the seam at overlap boundaries while preserving reconstruction accuracy in the overlap area."],"supporting_citations":[{"why":"Defines neural volume rendering and the photometric loss that SDF training and registration refinement build on.","marker":"[4]"},{"why":"Introduces the SDF-to-density weighting that lets a signed distance field be trained through volume rendering.","marker":"[5]"},{"why":"Provides the standard single-MLP SDF baseline compared in the experiments and the reconstruction model used at each node for Lego and Jade.","marker":"[6]"},{"why":"Supplies the high-capacity neural surface model used for the human-body, BlendedMVS, and campus reconstructions.","marker":"[8]"},{"why":"The structure-from-motion system that independently estimates camera poses for each node's image subset.","marker":"[9]"},{"why":"The block-based NeRF method whose divide-and-conquer idea is extended here from color radiance fields to SDF surfaces.","marker":"[12]"},{"why":"Shows that volume rendering can optimize camera parameters, the principle behind the registration refinement.","marker":"[45]"}],"fun_headline_variants":["Graph of local neural SDFs cuts reconstruction error by 45.6%","Scalable neural SDF graph: 45.6% lower error, 1 km campus demo","Divide-and-conquer neural SDFs: high quality, scalable 3D","Local neural SDFs beat single network: 45.6% error cut","Overlapping neural SDFs: 45.6% error drop, scale to 1 km scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that any two overlapping node reconstructions differ only by one global rotation, translation, and scale; if per-node camera estimation leaves non-rigid drift or locally varying scale, the registration has no model for it and the color-based refinement must silently compensate.","fun_headline_variants_meta":{"raw":{"variants":["Graph of local neural SDFs cuts reconstruction error by 45.6%","Scalable neural SDF graph: 45.6% lower error, 1 km campus demo","Divide-and-conquer neural SDFs: high quality, scalable 3D","Local neural SDFs beat single network: 45.6% error cut","Overlapping neural SDFs: 45.6% error drop, scale to 1 km scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000858,"raw_usage":{"total_tokens":3751,"prompt_tokens":996,"completion_tokens":2755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2650}},"tokens_in":612,"tokens_out":2755,"duration_ms":17859,"temperature":1.0,"reasoning_tokens":2650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:22:59.288507+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pair of overlapping nodes from a large scene, apply a smooth non-rigid warp (for example, a 5% scale gradient across the overlap) to one node's structure-from-motion output, then run the full registration; if the optimized color-consistency loss converges to a visible seam or a wrong transform, the single-similarity-transform assumption is falsified.","supporting_citations":[{"cited_title":"Nerf: Representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Defines neural volume rendering and the photometric loss that SDF training and registration refinement build on."},{"cited_title":"Volume rendering of neural implicit surfaces,","cited_arxiv_id":null,"evidence_quote":"Introduces the SDF-to-density weighting that lets a signed distance field be trained through volume rendering."},{"cited_title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the standard single-MLP SDF baseline compared in the experiments and the reconstruction model used at each node for Lego and Jade."},{"cited_title":"Neuralangelo: High-fidelity neural surface reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the high-capacity neural surface model used for the human-body, BlendedMVS, and campus reconstructions."},{"cited_title":"Structure-from-motion re- visited,","cited_arxiv_id":null,"evidence_quote":"The structure-from-motion system that independently estimates camera poses for each node's image subset."},{"cited_title":"Block-nerf: Scalable large scene neural view synthesis,","cited_arxiv_id":null,"evidence_quote":"The block-based NeRF method whose divide-and-conquer idea is extended here from color radiance fields to SDF surfaces."},{"cited_title":"iNeRF: Inverting neural radiance fields for pose estima- tion,","cited_arxiv_id":null,"evidence_quote":"Shows that volume rendering can optimize camera parameters, the principle behind the registration refinement."}],"review_version":1}