{"id":"7d4d6e46-5c89-42af-8c4e-431665abdb4d","arxiv_id":"2505.04668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":14,"one_line_summary":"SGCR reconstructs 3D parametric curves from multi-view images by optimizing Spherical Gaussians against 2D edge maps and then fitting rational Bezier curves.","lead":"This paper reconstructs 3D edge curves from ordinary multi-view photos by training small Gaussian spheres to match 2D edge maps, then linking the spheres into smooth Bezier curves. It reports higher accuracy and much lower training time than four prior curve-reconstruction systems on two benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's SOTA margins are small and unreported with uncertainty; without error bars or a held-out validation split, the claimed advantage over EMAP/NEF is not established.","rationale":"I read the paper in good faith. The method is coherent, the two-stage pipeline is clearly described, and the ablation study in Table 2 gives real evidence that each component matters, including the opacity-color loss Loc: removing it degrades CD from 0.028 to 0.051. This directly supports the reader's identified assumption that cross-view edge information plus Loc is sufficient to keep occluded or missed edges alive. So the reader's weakest_assumption is not the soft spot I would emphasize. The softer point is the experimental evidence for the headline 'state-of-the-art' claim. The differences over EMAP in Table 1 are small, no uncertainty is reported, and the optimization has stochastic elements whose seeds and settings are unspecified. Since the paper tunes hyperparameters on the same ABC-NEF benchmark used for the final comparison, the reported advantage may reflect configuration choices rather than a robust algorithmic gain. This is a correctness risk for the central claim, not a disagreement with the method's general approach. A simple significance test with fixed hyperparameters and multiple seeds would settle whether the concern lands. If the margins survive, the claim stands; if they do not, the verdict should remain conditional or move toward rejection of the specific SOTA claim. Because the reader already issued a CONDITIONAL verdict and my concern falls within the same conditions (error bars, validation separation, code release), I do not recommend changing the verdict, only sharpening the reason.","tokens_in":13022,"tokens_out":6239,"duration_ms":65676,"concrete_test":"Run SGCR 10 times with different random seeds on a fixed subset of 10 ABC-NEF scenes, and compute the mean and standard deviation of CD and F-score; do the same for NEF and EMAP using their released pretrained models and the same input edge maps. If the 95% confidence intervals overlap, or if EMAP's margin falls within SGCR's run-to-run spread, then Table 1's 'outperforms' claim is not statistically supported. In the same protocol, freeze all hyperparameters on a 20-model validation split before evaluating on the 115-model ABC-NEF test set; if the frozen-hyperparameter margin disappears, the headline result is an artifact of test-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SGCR outperforms existing methods rests on Table 1. The decisive comparisons are close: on ABC-NEF, CD is 0.0280 vs EMAP's 0.0291 (about 4% relative) and F-score is 0.926 vs 0.921; on ModelNet, CD is 0.0264 vs 0.0290 and F-score is 0.8952 vs 0.8784. No standard deviations, confidence intervals, per-scene breakdowns, or numbers of runs are reported. This matters because the pipeline is stochastic: Algorithm 1 repeatedly calls RandomChoice to seed line fits, and the number of random searches n, the interpolation count Ns, the optimizer settings, and random seeds are not specified. The quoted numbers could therefore be a favorable single run. In addition, hyperparameters such as r0=0.005, lambda1=0.2, lambda2=2, lambda3=0.01, delta1=0.02, delta2=0.01, and lambda=0.005 are selected using ablations on the same ABC-NEF benchmark that later produces the headline comparison, with no held-out validation split. A 4% CD improvement and a 0.5-point F-score improvement over EMAP can easily be within run-to-run variation or tuning noise. The efficiency claim (87 s vs over 1 h for NEF and over 2 h for EMAP) also lacks hardware and timing-protocol details, but the accuracy claim is the load-bearing part because the paper's title and abstract emphasize reconstruction quality as well as efficiency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SGCR, a two-stage method for reconstructing 3D parametric edge curves from calibrated multi-view images. In the first stage, the method optimizes a set of spherical Gaussians—isotropic, fixed-radius, grayscale primitives—using a view-based rendering loss that compares rendered 2D edge maps against PiDiNet-extracted edge maps, augmented by an opacity-color consistency loss and a regularization loss, within a two-phase training and pruning strategy. In the second stage, a RANSAC-style line fitting algorithm (Algorithm 1) followed by global optimization of rational Bézier control points and weights extracts the final curves, using a weighted Chamfer distance and an endpoint-connection loss. Experiments are reported on ABC-NEF, ModelNet, DTU, and Replica, claiming state-of-the-art accuracy on ABC-NEF and ModelNet, with training time of 87 seconds per scene versus over one hour for NEF and over two hours for EMAP.","tokens_in":13454,"tokens_out":5037,"duration_ms":49076,"significance":"If the results are robust, the paper offers a practical and interpretable alternative to neural implicit edge fields such as NEF and EMAP, with an order-of-magnitude training speedup and a simple geometric primitive that could be reused in other geometry-from-images tasks. The release of code and the systematic ablations of each pipeline component are strengths. However, the claimed accuracy advantage over EMAP is small (e.g., CD 0.0280 vs 0.0291, F-score 0.926 vs 0.921 on ABC-NEF), and the statistical evidence for this advantage is currently thin, which tempers the significance of the headline claim.","major_comments":[{"comment":"Table 1 reports improvements over EMAP that are small in absolute terms (CD 0.0280 vs 0.0291; F-score 0.926 vs 0.921 on ABC-NEF), but no standard deviations, confidence intervals, number of runs, or per-scene breakdowns are provided. The pipeline is stochastic: Algorithm 1 uses RandomChoice for seed line fits, and the Gaussian optimization presumably has random initialization and pruning order. Without multiple seeded runs and a significance test, the claimed superiority over EMAP is not established and could be within run-to-run variation.","section":"Table 1"},{"comment":"The central accuracy claim is based on ABC-NEF comparisons in Table 1, but the hyperparameters (r0, lambda1, lambda2, lambda3, delta1, delta2, lambda, and pruning thresholds) are selected by ablations run on the same ABC-NEF dataset, as shown in Tables 2 and 3. No held-out validation split is described. This creates a risk that the reported improvement is partly a result of tuning on the test set, and it weakens the generality of the results on other datasets. A proper validation split or cross-dataset hyperparameter transfer should be reported.","section":"Section 4.4 / Tables 2 and 3"},{"comment":"The input point cloud for the point-cloud baselines (RFEPS and NerVE) is not specified. If these methods are given ground-truth point clouds from the CAD models while NEF, EMAP, and the proposed method only receive rendered images, the comparison is not apples-to-apples and would understate the difficulty of the image-based setting. If the point clouds are derived from the same rendered images, the derivation procedure must be described. Please clarify this protocol.","section":"Section 4.2 / Comparisons"},{"comment":"Algorithm 1 depends on several parameters that are never specified: n (the number of random searches), Ns (the number of interpolation points), and the exact optimization procedure for p and q (e.g., optimizer, iterations, learning rate). These details are essential for reproducibility. Similarly, the Spherical Gaussian training stage omits optimizer settings, learning rate, densification thresholds, and the initialization scheme for the grid. Without these, an independent implementation cannot verify the reported numbers.","section":"Section 3.3.1 / Algorithm 1"}],"minor_comments":[{"comment":"The text says \"The regularization loss is defines as\" which should be \"is defined as\"; please also check the formula formatting for readability.","section":"Equation (4)"},{"comment":"The header contains the typo \"Sperical Gaussians\" which should be \"Spherical Gaussians\".","section":"Table 2"},{"comment":"The expression for L_endpoints is algebraically equivalent to a sum of squared endpoint distances for pairs within delta2, but the current form with the min multiplied by the fraction is confusing and appears to contain a division that cancels; please simplify for clarity.","section":"Equation (10)"},{"comment":"N0 is introduced in the text before Algorithm 1 but is not labeled inside the algorithm; adding a line defining N0 in the pseudocode would make the algorithm self-contained.","section":"Section 3.3.1"},{"comment":"The sentence \"which consisting of 115 distinct and challenging CAD models\" contains a grammatical error; it should be \"which consists of\".","section":"Section 4.1"},{"comment":"The real-world evaluation on DTU and Replica is only qualitative. Adding quantitative metrics, or at least a discussion of failure cases, would strengthen the claim of generalization.","section":"Section 4.3"},{"comment":"The phrase \"Spherical Gaussians serve as intermedia\" should be \"intermediate\".","section":"Abstract"},{"comment":"Please specify the GPU/CPU hardware used for all training and reconstruction time measurements, as the efficiency comparison is a central part of the paper's claim.","section":"Table 1 / Timing"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the code availability is a plus, but the evaluation falls short of supporting the strong state-of-the-art claim. The small margins over EMAP, the absence of error bars, and the tuning on the test benchmark are the main concerns. A revision that adds statistical rigor, clarifies the baseline protocols, and supplies full implementation details would make the paper acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth reading if you care about 3D reconstruction from images. It introduces Spherical Gaussians—isotropic, fixed-radius 3DGS primitives with gray color—and optimizes them against 2D edge maps with a view-consistency loss, then extracts rational Bezier curves via a two-stage line-fitting and global-optimization routine. The combination is genuinely new, and it works: on ABC-NEF and ModelNet it posts the best Chamfer distance and F-score among the compared methods, and training takes 87 seconds versus hours for NEF/EMAP. That efficiency claim is the strongest part of the paper; it is a large, qualitative gap, not a statistical one.\n\nThe paper also does a few things right: the ablation table is systematic and shows each component (grid init, edge loss, opacity-color loss, two-phase training, line fitting, global optimization) contributes. The real-world demos on DTU and Replica give some generalization signal. Code is linked, though not pinned to a commit.\n\nThe soft spots are real but not fatal. The reported SOTA margins over EMAP are small: on ABC-NEF, CD 0.0280 vs 0.0291, F-score 0.926 vs 0.921. No error bars, no per-scene breakdown, no number of runs. The pipeline is stochastic—Algorithm 1 does random line searches—so a single run is a plausible explanation for a 4% edge. Worse, several key hyperparameters (r0, lambda1-3, delta1-2) are chosen via ablations on ABC-NEF and then the headline numbers are reported on the same benchmark. That is tuning on the test set, and it weakens the accuracy claim more than the efficiency claim. The timing comparison also lacks hardware details, but that is minor.\n\nThe central idea is sound. The spherical Gaussian representation is a reasonable constraint that gives geometric meaning to Gaussian splats, and the curve extraction is a sensible way to turn discrete spheres into parametric curves. The ablation shows the isotropic constraint matters: replacing it with standard 3DGS collapses the quality (CD 0.142). So the paper deserves a serious referee, but the revision must add uncertainty quantification, a held-out hyperparameter split (or at least an honest statement that the reported numbers are on the tuning set), and details on the random search and timing protocol.\n\nWho is this for? Anyone working on edge/curve reconstruction from multi-view images, or on constrained Gaussian representations for geometry. It is a solid subfield contribution, not a paradigm shift. I'd accept it for review and would cite it if I were working on curve reconstruction.\n\nMy take: accept with major revision, pending error bars and a cleaner evaluation protocol.","headline":"A well-engineered pipeline that makes 3D curve reconstruction from images fast; the SOTA accuracy claim is real but rests on small margins without error bars.","tokens_in":13943,"tokens_out":2060,"would_cite":true,"duration_ms":19889,"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":"Fixed-radius spherical Gaussians turn multi-view 2D edge maps into accurate 3D parametric curves, cutting training from hours to under two minutes.","keywords":["3D curve reconstruction","Spherical Gaussians","3D Gaussian Splatting","multi-view reconstruction","edge detection","parametric curves","rational Bezier curves","neural rendering"],"falsifier":"Take a multi-view set in which one sharp edge is occluded in all but a single view, run the pipeline, and compare the reconstructed curve to a ground-truth scan: if the opacity-color loss cannot keep the occluded Gaussians alive, the edge will be absent from the wireframe and the Chamfer distance to ground truth will jump above the matching threshold.","tokens_in":12868,"feed_emoji":"📐","tokens_out":5098,"duration_ms":49679,"temperature":0.7,"pith_summary":"This paper tries to establish that 3D feature curves can be reconstructed directly from calibrated multi-view images by replacing standard anisotropic 3D Gaussians with fixed-radius isotropic spheres, called Spherical Gaussians, and optimizing them against 2D edge maps alone. The authors claim that these spheres align with true 3D edges under a view-based rendering loss plus an opacity-color consistency term, and that a subsequent optimization-based extraction, SGCR, turns the spheres into smooth rational Bezier curves. If correct, accurate parametric edge reconstruction no longer needs point clouds, 3D supervision, or hours of neural-field training: the reported pipeline trains in about a minute and extracts curves in about half a minute per scene.","feed_headline":"Spherical Gaussians rebuild 3D edges from photos in 87 seconds","feed_subtitle":"A 2D-supervised pipeline beats prior curve reconstruction on benchmarks while cutting single-scene training from hours to minutes.","key_machinery":"The central object is the Spherical Gaussian: a Gaussian primitive with no covariance matrix and no spherical harmonics, defined only by center, fixed radius, opacity, and gray color. Its role is to act as a small isotropic atom that a rendering loss can push onto 3D edges while remaining geometrically interpretable. The opacity-color loss $L_{\\mathrm{oc}}$ couples each sphere's opacity to its color so that occluded or detector-missed edges are not pruned prematurely. The extraction algorithm SGCR then uses line fitting with dilated Chamfer distance followed by global rational Bezier optimization to produce parametric curves.","core_discovery":"Spherical Gaussians are 3D Gaussian primitives stripped of covariance and spherical harmonics: each is a sphere of fixed radius carrying only opacity and gray-scale color. The paper claims that when initialized on a regular grid and trained with an edge-aware rendering loss, an opacity-color consistency loss, opacity regularization, and two-phase pruning, these spheres settle on the sharp edges of the object rather than forming the redundant ellipsoids produced by ordinary 3D Gaussian Splatting. The SGCR extraction then fits third-order rational Bezier curves through the sphere centers, first by greedy line fitting with dilated Chamfer distance and then by global optimization with opacity-weighted Chamfer distance and endpoint-joining loss. On the ABC-NEF and ModelNet benchmarks the method is reported to beat prior edge-reconstruction methods on Chamfer distance and F-score, while cutting single-scene training from over an hour to 87 seconds.","pith_inferences":["A stronger 2D edge detector should directly improve the 3D curves, since the method inherits the detector's misses and false positives; this suggests a modular pipeline where detector and curve extraction can be improved independently.","The same fixed-radius isotropic primitive could be extended to other feature types, such as corners, creases, or semantic curves, by changing only the rendering target.","The method's speed opens the door to online reconstruction from video streams, where each frame updates the spheres rather than restarting training."],"forward_implications":["3D curve reconstruction from images drops from hours to minutes per scene, making interactive or near-real-time pipelines feasible.","Because supervision is only 2D edge maps, the method scales to arbitrary captured image sets without 3D annotations or point clouds.","The extracted output is parametric rational Bezier curves, directly usable in CAD, reverse engineering, and surface-reconstruction downstream tasks.","The sphere representation may serve as a general bridge between 2D rendering losses and 3D line-like geometry, beyond edges."],"supporting_citations":[{"why":"Supplies the base Gaussian Splatting representation and rendering pipeline that Spherical Gaussians simplify and retarget to edge maps.","marker":"[19]"},{"why":"Provides the pretrained 2D edge detector that produces the ground-truth edge maps used as supervision.","marker":"[38]"},{"why":"Defines the ABC-NEF benchmark and the neural edge field baseline that the method compares against and improves on.","marker":"[50]"},{"why":"Serves as the other main image-based baseline for 3D edge reconstruction and efficiency comparison.","marker":"[23]"},{"why":"One of the point-cloud-input baselines used for quantitative and qualitative comparison.","marker":"[47]"},{"why":"The other point-cloud baseline, included to show the method works without 3D input.","marker":"[54]"},{"why":"Provides the CAD models with edge annotations used to build the evaluation data.","marker":"[20]"}],"fun_headline_variants":["Spherical Gaussians turn photos into 3D curves in 87 seconds","Spherical Gaussians make 3D curve reconstruction fast and accurate","Accurate 3D curves from photos in 87 seconds with Spherical Gaussians","Spherical Gaussians beat prior 3D edge methods in speed and accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that 2D edge maps, together with the opacity-color consistency loss, contain enough cross-view information to place spheres on true 3D edges even when those edges are occluded or missed in some views.","fun_headline_variants_meta":{"raw":{"variants":["Spherical Gaussians turn photos into 3D curves in 87 seconds","Spherical Gaussians make 3D curve reconstruction fast and accurate","Accurate 3D curves from photos in 87 seconds with Spherical Gaussians","Spherical Gaussians beat prior 3D edge methods in speed and accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3558,"prompt_tokens":963,"completion_tokens":2595,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":2512}},"tokens_in":579,"tokens_out":2595,"duration_ms":16665,"temperature":1.0,"reasoning_tokens":2512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:34:23.620573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a multi-view set in which one sharp edge is occluded in all but a single view, run the pipeline, and compare the reconstructed curve to a ground-truth scan: if the opacity-color loss cannot keep the occluded Gaussians alive, the edge will be absent from the wireframe and the Chamfer distance to ground truth will jump above the matching threshold.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the base Gaussian Splatting representation and rendering pipeline that Spherical Gaussians simplify and retarget to edge maps."},{"cited_title":"Pixel difference networks for efficient edge detection","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained 2D edge detector that produces the ground-truth edge maps used as supervision."},{"cited_title":"Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images","cited_arxiv_id":null,"evidence_quote":"Defines the ABC-NEF benchmark and the neural edge field baseline that the method compares against and improves on."},{"cited_title":"3d neural edge recon- struction","cited_arxiv_id":null,"evidence_quote":"Serves as the other main image-based baseline for 3D edge reconstruction and efficiency comparison."},{"cited_title":"Rfeps: Re- constructing feature-line equipped polygonal surface","cited_arxiv_id":null,"evidence_quote":"One of the point-cloud-input baselines used for quantitative and qualitative comparison."},{"cited_title":"Nerve: Neural volumetric edges for parametric curve extraction from point cloud","cited_arxiv_id":null,"evidence_quote":"The other point-cloud baseline, included to show the method works without 3D input."},{"cited_title":"Abc: A big cad model dataset for geometric deep learning","cited_arxiv_id":null,"evidence_quote":"Provides the CAD models with edge annotations used to build the evaluation data."}],"review_version":1}