{"id":"8ee5f493-63a8-4132-b7ac-c2ede2975406","arxiv_id":"2411.17580","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A new real-world railway point cloud completion dataset shows existing methods fail on noisy, non-uniform scans, and a proposed 'homology sampler' network improves results.","lead":"RealPC is a new paired dataset of about 40,000 incomplete and complete 3D point clouds of railway structures, captured from real scans rather than clean CAD models. Popular completion models perform much worse on it, and the paper proposes a network that uses shape-skeleton-like priors to improve completion.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"0-dimensional PH cannot define a 3D skeleton, and Eq. 1 is internally inconsistent; the paper's central claim that PH priors improve completion is unsupported.","rationale":"The reader's weakest assumption correctly identifies the H0-to-skeleton mapping as the load-bearing gap: the paper needs 0-dim PH to provide a spatial backbone, but standard PH provides only merge scales. I add two concrete corroborating problems that make the gap decisive rather than merely underjustified. First, Eq. 1 is algebraically reversed for H0 features: with birth values equal to zero, minimizing sum(b_i - d_i) maximizes persistence, not minimizes it, so the loss as printed cannot implement the stated connectivity objective. Second, BOSHNet's reported success is confounded because Eq. 2 feeds the network sparse samples of the complete ground-truth cloud during training; the method is therefore not a clean test of whether topological priors help, since a nontopological baseline that trains on arbitrary GT surface subsamples would likely produce the same or better numbers. These issues all bear on the same central claim: that integrating PH priors into existing completion models improves completion. If that claim were true, the paper would need a correct loss, a derivation showing what H0 actually controls, and an ablation isolating the topological component from the GT-subsampling component. None is present. The dataset contribution and the comparative PH analysis of RealPC may still be useful, but the methodological claim as stated is not established, so the reader's REJECT verdict stands unchanged.","tokens_in":19345,"tokens_out":6733,"duration_ms":69101,"concrete_test":"Take a straight-line point cloud and a Y-shaped point cloud with equal point counts and identical pairwise merge scales (for example, by arranging the MST edge lengths of both shapes to match). Compute their 0-dim persistence diagrams. If the diagrams are identical, H0 features cannot encode the skeleton geometry asserted in Section 5.1.2 and Figure 7, which would directly falsify the claimed PH-prior mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Section 5.1.2's assertion that 0-dim PH extracts a complete PC skeleton. This is not what H0 persistence provides. Each H0 birth-death pair records only the scale at which a connected component merges; it carries no spatial coordinates, no edges, and no branch/backbone structure. Minimizing H0 persistence is equivalent to minimizing the total length of a minimum spanning tree of the points: it makes the seed cloud compact and connected, but it does not place generated points along any particular geometric skeleton. Eq. 1 as printed compounds the problem: it minimizes sum(b_i - d_i). For a Vietoris-Rips filtration on a point cloud, every H0 component is born at b=0, so b_i - d_i is minus the persistence; minimizing this sum maximizes death times, i.e., maximizes persistence, the opposite of the stated goal of merging components early. The visual skeleton in Fig. 7 is therefore not an output of the loss as written. BOSHNet's Eq. 2 is also confounded: its first term trains Net on BOSH(c_i,j), sparse samples of the complete ground-truth cloud, so the large CD-L2 improvement in Table 5 can be explained by learning to copy ground-truth surface samples rather than by any topological prior. No ablation controls for this. Without a valid derivation connecting H0 to a skeleton and without an ablation that removes the GT-subsampling confound, the paper's proof claim is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript has two main goals. First, it introduces RealPC, a paired real-world industrial point cloud completion dataset with roughly 40,000 partial/complete pairs across 21 railway-related object categories, built from four open scene-level LiDAR datasets. The paper characterizes RealPC against PCN and ShapeNet using noise, non-uniformity, and 0- and 1-dimensional persistent homology (H0/H1) statistics, and benchmarks several completion, reconstruction, and generation baselines on it. Second, it proposes topological-prior methods: TopODGNet, which adds a 0-dim PH regularizer to ODGNet, and BOSHNet, which replaces homology computation with a 'Homology Sampler' that samples proxy backbones from ground-truth surfaces. The paper claims to prove that integrating these topological priors improves completion performance, citing Table 5 as evidence.","tokens_in":19680,"tokens_out":7959,"duration_ms":114403,"significance":"If the claims were correct, the paper would make a significant contribution: a real-world paired object completion benchmark, a systematic failure analysis of synthetic-trained models on real data, and a principled topological regularizer. The dataset itself is potentially valuable, and the comparative noise/non-uniformity/PH analysis is suggestive and useful. However, the central methodological claim is not currently supported. The topological loss in Eq. (1) is mis-specified relative to its stated objective, the claimed relationship between H0 persistence and a geometric skeleton is not established, BOSHNet's training loss leaks ground-truth samples, and the headline experimental table is internally inconsistent. These are load-bearing issues for the paper's main claim, not presentation problems.","major_comments":[{"comment":"The topological loss as written, sum over i>k of (b_i - d_i), is inconsistent with the stated goal. In a Vietoris–Rips filtration, every 0-dimensional homology class is born at b_i = 0, so the summand equals -d_i and minimizing the loss maximizes death times, i.e., maximizes persistence. This is the opposite of the text's claim that the loss minimizes persistence so that components merge early. The paper also switches conventions: Section 3 defines persistence as death minus birth, while Section 5.1.2 calls (b-d) the persistence. As a result, the mechanism claimed for TopODGNet is not the mechanism implemented by the printed loss, and the abstract's 'proof' claim is not supported by the actual equation.","section":"Section 5.1.2, Eq. (1)"},{"comment":"The paper asserts that '0-dim PH ensures the extraction of a complete PC skeleton' and that minimizing H0 persistence guides point generation along that skeleton. H0 birth-death pairs only record the scales at which connected components merge; they carry no spatial coordinates, no edges, and no branch or backbone structure. The manuscript provides no algorithm or derivation that converts the H0 persistence diagram into the rendered 'skeleton' of Figure 7. At best, an H0 loss encourages the seed cloud to become connected or compact; it does not place generated points on a particular geometric skeleton. The explanatory mechanism for TopODGNet is therefore unsupported, and BOSHNet's sampled subsets are not demonstrated to be proxies for H0 persistence in any formal sense.","section":"Section 5.1.2 and Figure 7"},{"comment":"The first term of Eq. (2) is M(Net(BOSH(c_i,j)), c_i), where c_i is the complete ground-truth cloud and BOSH(c_i,j) are sparse subsets sampled from that same complete cloud. This trains the network to reconstruct a complete cloud from ground-truth-derived samples, which is a much easier task than completion from a real partial scan and is unavailable at test time. The large CD-L2 improvement reported for BOSHNet in Table 5 could therefore be an artifact of this ground-truth supervision shortcut rather than evidence for a topological prior. The paper needs ablations that (a) train the same backbone with random or partial-input samples in place of BOSH samples, and (b) compare against an oracle that simply upsamples the BOSH samples. Without these, the superiority claim for BOSHNet is not established.","section":"Section 5.2, Eq. (2)"},{"comment":"The reported quantitative evidence is internally inconsistent. Under the metric definitions in Section 9, CD-L2 is a sum of squared nearest-neighbor distances, so for point clouds with nearest-neighbor distances below 1 it must be smaller than CD-L1. Yet Table 5 reports SnowflakeNet with CD-L1=60 and CD-L2=72, while BOSHNet is reported with CD-L1=69 and CD-L2=5.4. The drastically different relationship between CD-L1 and CD-L2 across rows is not explained, and no error bars or repeated runs are provided for any method. This undermines the quantitative basis for the central claim that topological priors improve completion.","section":"Table 5 and Section 6"}],"minor_comments":[{"comment":"The words 'prove' and 'proof' are used in the abstract and contributions, but the paper contains no formal theorem or rigorous mathematical proof; recommend replacing these with 'demonstrate' or 'show'.","section":"Abstract and Contributions"},{"comment":"Please specify exactly how the 'average of persistence averages' is computed across classes, and how the essential 0-dimensional class of infinite persistence is treated when averaging H0 persistence values.","section":"Section 4.2"},{"comment":"The SnowflakeNet row contains missing entries represented by '-' without a note; please state how the class mean was computed from the available classes, or report complete results.","section":"Table 4"},{"comment":"The statement that setting k=1 'works fairly well' appears to conflict with the main text's discussion that k>=2 is needed for disconnected partial inputs; please clarify the criterion and sensitivity for selecting k.","section":"Supplementary Section 12.1"},{"comment":"Please specify what is plotted as the 'skeleton' and how it is obtained from the persistence pairs; without this, the figure cannot be reproduced.","section":"Figure 7"},{"comment":"The dataset release statement says the data and code will be made open-source, but no URL or license is provided; please include the actual artifact links.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The RealPC dataset itself could be a publishable contribution if separated from the topological-prior method claims. As submitted, however, the central claim is unsupported by load-bearing issues in the loss formulation, the skeleton interpretation, the BOSHNet training setup, and the headline results. These are not merely local fixes; they require rethinking the method and rerunning the experiments. I would encourage the authors to resubmit a version focused on the dataset and benchmark, with the method claims substantially revised or removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe RealPC dataset is the real news here; the method claims are not. The paper contributes a paired real-world point cloud completion benchmark built from railway scene scans, ~40k partial-complete pairs across 21 categories, and shows convincingly that standard baselines degrade sharply on it. That addresses a genuine gap: completion models are almost exclusively trained and tested on synthetic CAD-derived clouds, and a real-world benchmark with this scale could shift evaluation practice. The comparative persistent homology analysis, showing that RealPC clouds have richer H0/H1 structure than PCN/ShapeNet, is a nice descriptive use of TDA, though unnormalized and only suggestive.\n\nThe methodological core does not hold up. The claim that 0-dim PH extracts a 3D skeleton is not what H0 provides: birth-death pairs record merge scales, not spatial coordinates or branch structure. Minimizing H0 persistence, at best, makes the seed cloud compact and connected; it does not generate points along a geometric skeleton. Worse, Eq. (1) as printed minimizes sum(b_i - d_i). For a Vietoris–Rips filtration on a point cloud, every H0 feature is born at 0, so b_i - d_i = -d_i, and minimizing this sum maximizes death times — the opposite of the stated goal of merging components early. The sign error is load-bearing. The visual skeleton in Fig. 7 is not an output of the loss as written.\n\nBOSHNet's Eq. (2) compounds the problem: the first term trains the network on sparse samples of the complete ground-truth cloud and asks it to reconstruct that same cloud. That is an oracle-subsampling shortcut, not a topological prior. The large CD-L2 improvement over baselines in Table 5 can be explained by the network learning to copy surface samples, and there is no ablation that removes this confound. Table 5 also has an internal metric inconsistency: for SnowflakeNet, CD-L2 (72) exceeds CD-L1 (60), which clashes with the CD-L2/CD-L1 relationship one would expect for normalized clouds, and with BOSHNet's own numbers (CD-L1 69, CD-L2 5.4). That signals sloppy evaluation.\n\nThe dataset, if released with the processing pipeline, could be a useful resource. The paper as submitted overreaches with its \"proof\" that topological priors improve completion; that part needs either a correct derivation and controlled ablations, or to be removed. The benchmark and topological comparison deserve referee attention, but the method section needs substantial revision before the paper is acceptable.\n\nSend it to peer review — a serious editor should not desk-reject, since the dataset is a real contribution — but the reviewers should be told to focus on the method section's load-bearing flaws.","headline":"RealPC is a genuinely useful real-world completion benchmark, but the paper's topological-prior method is unsupported and needs a major rework.","tokens_in":20212,"tokens_out":4960,"would_cite":false,"duration_ms":42475,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding 0-dimensional persistent-homology priors to point cloud completion networks improves their outputs on real-world scans, and a new dataset, RealPC, shows why synthetic-trained models fail.","keywords":["point cloud completion","RealPC dataset","persistent homology","0-dimensional homology priors","Homology Sampler","topological data analysis","railway infrastructure point clouds","Chamfer distance"],"falsifier":"Train TopODGNet with the topology loss replaced by a control loss that merely minimizes the sum of pairwise distances among seed points, without any persistence computation; if the Chamfer-distance improvement over ODGNet matches TopODGNet's, then the 0-dim PH skeleton prior is not the active ingredient, and the authors' topological explanation collapses.","tokens_in":19119,"feed_emoji":"🏗️","tokens_out":9070,"duration_ms":80098,"temperature":0.7,"pith_summary":"Point cloud completion models are typically trained and evaluated on clean synthetic objects, and this paper argues that those benchmarks hide what real-world scans look like: noisy, non-uniformly dense, and full of topological structure. To make the case, the authors build RealPC, roughly 40,000 partial-to-complete pairs of industrial railway structures extracted from four scene-level LiDAR datasets, and show that strong completion networks that look good on PCN collapse on RealPC. Using persistent homology, they find that RealPC point clouds have many significant 0- and 1-dimensional topological features, unlike synthetic clouds, and they claim that feeding those features back as priors improves completion: a topology-loss variant of ODGNet reduces Chamfer distance, and BOSHNet, which samples cheap proxy backbones from the complete surface instead of computing homology, reaches the best reported numbers. The paper's central proposal is that topological consistency, not just pointwise closeness, is a missing ingredient for completing real-world shapes.","feed_headline":"Topology priors improve point cloud completion on real-world scans","feed_subtitle":"New RealPC benchmark shows why synthetic-trained models fail, and how topological backbones fix it.","key_machinery":"The load-bearing mechanism is the 0-dim PH backbone: 0-dimensional persistent homology records, for each connected component born during a distance filtration, the radius at which it merges, and the paper interprets these birth-death pairs as outlining a spatial skeleton of the complete point cloud. The topology loss $L_{topo} = \\sum_{i>k}(b_i - d_i)$ minimizes the persistence of those pairs on the sparse decoder seeds, forcing the seeds to coalesce into one connected skeleton (or $k$ skeletons for multi-component inputs). Because computing Vietoris-Rips persistence is expensive, BOSHNet replaces the homology computation with a Homology Sampler that directly draws multiple sparse proxy backbones from the surface of the ground-truth complete cloud, and trains the network to complete those backbones as well as the partial input. The dataset RealPC itself is the other load-bearing piece: HDBSCAN clustering of four scene-level railway point-cloud datasets, followed by manual inspection for complete ground truth and three sparsification/incompletion processes (viewpoint-based farthest-point removal, viewpoint-based probability sampling, and uniform subsampling), is what produces paired real-world training data with measurable 0- and 1-dimensional persistence.","core_discovery":"The paper's claim, stated on its own terms, is that 0-dimensional persistent homology priors give a completion network a global skeleton of the complete shape, and that enforcing this skeleton during training produces more topologically consistent completions. The evidence is: on RealPC, ODGNet's Chamfer distance goes from 119 to 103 (CD-L1) and from 111 to 80 (CD-L2) when the decoder seeds are regularized by the 0-dim PH persistence loss; and BOSHNet, which bypasses homology by sampling multiple sparse proxy backbones from the ground-truth surface, reaches CD-L1 69 and CD-L2 5.4, improving on all baselines. The authors also report that 0-dim PH priors improve scene completion and reconstruction on KITTI. The paper frames this as proof that existing completion methods fail on real-world data because they ignore the topological richness that real scans contain.","pith_inferences":["If the skeleton interpretation is right, then completion models without explicit decoder seeds could be adapted by adding a lightweight auxiliary head that outputs a sparse skeleton and applying the same persistence loss, extending the recipe beyond ODGNet.","The BOSH proxy suggests a general training strategy: sample multi-scale skeleton-like subsets from ground truth as auxiliary completion targets, which is testable on any paired dataset and may act as a curriculum where the model first learns coarse structure.","The dataset-creation pipeline (clustering, manual ground-truth selection, and multiple corruption processes) could be rerun on other scene-level LiDAR collections, producing topological completion benchmarks for urban, agricultural, or indoor environments without new sensors.","A discriminating test: train BOSHNet on PCN or ShapeNet; if its advantage shrinks or vanishes on synthetic shapes, that would confirm topological richness, not the sampling scheme itself, is what drives the RealPC gains."],"forward_implications":["Benchmarking on RealPC rather than PCN or ShapeNet changes which methods look strong, and several baselines that report low error on synthetic data exceed 100 in CD-L1 on RealPC.","Existing seed-based completion models can be upgraded with a 0-dim PH regularization layer without architectural changes beyond accessing decoder seeds, as TopODGNet does.","Homology computation can be bypassed: proxy backbones sampled from the complete surface give most of the benefit at far lower cost, and they are available from the first training step rather than emerging late in training.","Topological priors transfer beyond objects: KITTI scene completion and reconstruction improve with the same 0-dim PH loss.","RealPC supplies paired partial/complete real-world data, enabling supervised training that datasets like ScanNet, Matterport3D, and KITTI cannot support directly."],"supporting_citations":[{"why":"It supplies the synthetic partial/complete pairs and the PCN baseline that the paper compares against on RealPC.","marker":"[64]"},{"why":"It provides the seed-decoder completion model that TopODGNet builds on and the primary baseline in Table 5.","marker":"[6]"},{"why":"It is the hierarchical clustering algorithm used to separate individual structures from scene-level railway point clouds during RealPC creation.","marker":"[7]"},{"why":"It is the source CAD repository from which the PCN and ShapeNet comparison datasets are derived.","marker":"[11]"},{"why":"It is the real-world scene dataset used in the supplementary to test 0-dim PH priors on scene completion and reconstruction.","marker":"[25]"},{"why":"It supplies the simplification method used in the non-neural benchmark showing RealPC is harder to process than ShapeNet.","marker":"[31]"},{"why":"It provides the alpha-shape surface reconstruction method used to benchmark RealPC versus ShapeNet on mesh fidelity.","marker":"[20]"},{"why":"It is the diffusion-based shape generation model used to benchmark RealPC on generation quality.","marker":"[50]"},{"why":"It is the DGCNN backbone used for the KITTI scene experiments with and without PH priors.","marker":"[53]"}],"fun_headline_variants":["Real-world scans reveal completion models need topology priors","New dataset RealPC proves topology priors fix point completion","Topology priors: the missing key for real-world point cloud completion","How topology priors rescue point cloud completion on real data","RealPC benchmark: Existing completion methods fail, topology priors succeed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on identifying 0-dimensional persistent homology's merge-scale pairs with a spatial skeleton of the complete shape; standard 0-dim PH only records when connected components merge, and the paper assumes this defines a geometric backbone that guides generation.","fun_headline_variants_meta":{"raw":{"variants":["Real-world scans reveal completion models need topology priors","New dataset RealPC proves topology priors fix point completion","Topology priors: the missing key for real-world point cloud completion","How topology priors rescue point cloud completion on real data","RealPC benchmark: Existing completion methods fail, topology priors succeed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001082,"raw_usage":{"total_tokens":4556,"prompt_tokens":1010,"completion_tokens":3546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":3461}},"tokens_in":626,"tokens_out":3546,"duration_ms":41459,"temperature":1.0,"reasoning_tokens":3461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:58:46.234934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train TopODGNet with the topology loss replaced by a control loss that merely minimizes the sum of pairwise distances among seed points, without any persistence computation; if the Chamfer-distance improvement over ODGNet matches TopODGNet's, then the 0-dim PH skeleton prior is not the active ingredient, and the authors' topological explanation collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the synthetic partial/complete pairs and the PCN baseline that the paper compares against on RealPC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the seed-decoder completion model that TopODGNet builds on and the primary baseline in Table 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the hierarchical clustering algorithm used to separate individual structures from scene-level railway point clouds during RealPC creation."},{"cited_title":"Geiger, P","cited_arxiv_id":null,"evidence_quote":"It is the real-world scene dataset used in the supplementary to test 0-dim PH priors on scene completion and reconstruction."},{"cited_title":"Huang, D","cited_arxiv_id":null,"evidence_quote":"It supplies the simplification method used in the non-neural benchmark showing RealPC is harder to process than ShapeNet."},{"cited_title":"Edelsbrunner, D","cited_arxiv_id":null,"evidence_quote":"It provides the alpha-shape surface reconstruction method used to benchmark RealPC versus ShapeNet on mesh fidelity."},{"cited_title":"Vahdat, F","cited_arxiv_id":null,"evidence_quote":"It is the diffusion-based shape generation model used to benchmark RealPC on generation quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the DGCNN backbone used for the KITTI scene experiments with and without PH priors."}],"review_version":1}