{"id":"87305559-8572-4ab4-a4af-3e118aac44f4","arxiv_id":"2412.08421","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PointCFormer achieves state-of-the-art point cloud completion by weighting each neighbor's contribution with learned spatial and feature-space relations inside a progressive transformer extractor.","lead":"PointCFormer is a new neural network for 3D point cloud completion that combines learned neighbor-relation weights with transformer attention to preserve global shape and local detail. The paper reports state-of-the-art scores on four standard benchmarks, with small but consistent Chamfer distance improvements over prior methods such as AdaPoinTr.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The most load-bearing issue is Table 4's KITTI Fidelity value (0.001 vs 0.237 for AdaPoinTr), a ~200x outlier that is not plausible under the stated evaluation protocol; if it is a metric bug or typo, the claimed SOTA on KITTI is unsupported.","rationale":"I read the paper in good faith as presenting a coherent architecture with consistent improvements on PCN and ShapeNet-55, and the ablation study supports the role of the proposed local relation module. The central empirical claim is state-of-the-art performance on several benchmarks, so the load-bearing question is whether the reported benchmark numbers are trustworthy. The reader's weakest assumption focuses on unreported values M and lambda; those are real reporting gaps but not decisive, because the released code can supply the values and the relation metric's sensitivity to M is likely modest. The KITTI Fidelity number is a stronger internal red flag: it is an extreme outlier relative to all baselines, and it is not accompanied by any protocol detail that would explain how a real-scene completion can be essentially identical to the ground-truth scan at every generated point. This concern does not overturn the paper: even if Table 4 is wrong, the PCN and ShapeNet-55 results may still hold. It does, however, justify keeping the verdict conditional, and it identifies a specific check that should be run before the SOTA claim is accepted at face value.","tokens_in":15397,"tokens_out":10033,"duration_ms":111382,"concrete_test":"Download the released checkpoint from the provided GitHub repository and run the official KITTI evaluation code used by GRNet/AdaPoinTr to recompute Fidelity and MMD under the same protocol. If the recomputed Fidelity is materially above 0.001 (e.g., on the order of 0.1-0.3 like all baselines), Table 4 contains an error and the KITTI SOTA claim should be corrected. If the value 0.001 reproduces, report the exact evaluation protocol and explain how a dense completion can achieve near-zero distance to a partial LiDAR scan; if the explanation is not geometrically convincing, the metric computation should be examined.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is the KITTI Fidelity number in Table 4. PointCFormer is reported at 0.001, while the best baseline AdaPoinTr is 0.237, a roughly 200-fold improvement, yet the same table's MMD improvement is only from 0.392 to 0.353. Since the table header says '*1000', the PointCFormer value is 1e-6 under the usual interpretation; that would mean every generated point in the 16,384-point completion coincides with a ground-truth LiDAR point. On a real-scan benchmark where the ground truth is itself incomplete and noisy, this is not plausible without a different evaluation protocol, a metric bug, or a typo. If the value is wrong, the paper's claim to state-of-the-art performance on KITTI, one of the 'four common datasets' named in the contributions, is unsupported, and the trustworthiness of the other benchmark tables is cast into doubt because the same evaluation pipeline produced this outlier. The unreported M and lambda identified by the reader are legitimate but less decisive, since the released code can supply those values; this KITTI number is already in the paper and is internally suspect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents PointCFormer, a transformer-based network for point cloud completion. It introduces three modules: a local geometric relationship perception (LGRP) module that weights kNN neighbors via R1 (spatial Manhattan distance) and R2 (relative directed-edge deviation) metrics; a progressive feature extractor (PFE) that alternates self-attention and scale-tailored local feature extraction (ST-LFE) while downsampling; and a point proxy correction module (CM) that re-injects dense input information into latent point proxies. The method is trained on PCN, ShapeNet-55, ShapeNet-34/21, and KITTI, and the main tables report state-of-the-art Chamfer distances (CD-l1 6.41 on PCN, CD-l2 0.73 on ShapeNet-55) and favorable F-scores compared with AdaPoinTr, 3DMambaComplete, and others. An ablation on PCN shows that each module contributes to the final score.","tokens_in":15658,"tokens_out":5456,"duration_ms":49810,"significance":"If the reported results are accurate, the paper offers a practically useful architecture that improves completion quality across multiple benchmarks by a small but consistent margin (e.g., CD-l1 6.41 vs 6.53 for AdaPoinTr on PCN), and it provides a thorough module analysis and released code. The relation-based weighting idea is well motivated, and the progressive extractor's combination of global and local queries is a reasonable design. However, the empirical claim currently rests on at least one implausible result — the KITTI Fidelity of 0.001 — and on two unreported hyperparameters (M and lambda), which makes the significance conditional on those points being resolved.","major_comments":[{"comment":"Table 4 reports a KITTI Fidelity of 0.001 for PointCFormer versus 0.237 for AdaPoinTr, under a header stating values are multiplied by 1000. This implies an actual fidelity around 1e-6, i.e., nearly every completed point coincides with a ground-truth LiDAR point. Given the KITTI ground truth is itself incomplete and noisy, and the MMD improvement is only from 0.392 to 0.353, this value is not plausible under the stated protocol. The paper provides no explanation of any altered evaluation protocol. If this number is a typo or a unit error, the KITTI state-of-the-art claim is unsupported and the reliability of the other benchmark tables is cast in doubt because they are produced by the same pipeline. Please re-check the KITTI evaluation and report the corrected value or the exact protocol.","section":"Table 4 / Experiments"},{"comment":"The definition of R2 in Eq. (2) depends on a subset size M ('we select a small subset (M points) of the nearest neighbors'), and the final objective J_PC = J0 + J1 + λ Jdenoise depends on the weight λ. Neither M nor λ is reported in the main text or in the supplementary 'Hyper-parameter Recommendation' section, which states k=16 but is silent on M and λ. Without these values, the method cannot be reproduced, and the reported improvements could have resulted from fine-tuning M and λ on the test benchmarks. Please report the chosen values and provide a sensitivity analysis over reasonable ranges.","section":"Eq. (2) and Network Optimization"},{"comment":"The PointAttN(aaai2024) row in Table 1 is incomplete: values for Sofa, Table, Boat, and F-Score@1% are missing, and the Cabinet value 17.923 is far outside the range of all other methods (the other methods range from 8.34 to 10.47 on that category, except for PCN at 22.70). This incomplete row makes the 'leads in almost all metrics' claim hard to verify. Please provide the full PointAttN results as published in its original paper or clearly state why some entries are unavailable.","section":"Table 1"}],"minor_comments":[{"comment":"The supplementary material begins with 'Anonymous submission' and contains placeholder citations such as '(?)' in 'Similar to the AdapoinTr (?) protocol' and 'Projected-ShapeNet-55 (?)'. These need to be cleaned up before publication.","section":"Supplementary Material"},{"comment":"The main text cites 'Zhong et al. 2024' (the same arXiv preprint) for a more detailed analysis of R1 and R2, but the supplementary material appears to be part of this submission, not an external reference. This self-referential citation is confusing and should be replaced with a cross-reference to the supplement within this paper.","section":"References / Supplementary"},{"comment":"The supplementary tables report 'PointCFormer-Plus' results without any definition of this variant. If Plus is a larger model or an ensemble, define it in the main paper or remove the entries.","section":"Supplementary Tables 4-5"},{"comment":"The paper reports a single run for each method and no error bars or multiple seeds. Given that the reported margins over the closest competitor are sometimes small (e.g., 0.73 vs 0.77 CD-l2 on ShapeNet-55), a statement on variance or a multi-seed evaluation would strengthen the SOTA claim.","section":"Experiments"},{"comment":"In Table 2, the F-Score@1% of PointCFormer (0.499) is slightly lower than AdaPoinTr (0.503); the text acknowledges this but could state it more explicitly to avoid overclaiming.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The KITTI Fidelity number is the single most serious issue. If it cannot be corrected to a plausible value, the paper's claim of SOTA on KITTI should be dropped or the evaluation rerun. I still believe the method's core idea is reasonable and the other benchmark results look consistent with the ablations, so a major revision addressing the reported numbers, missing hyperparameters, and incomplete baselines is the appropriate path rather than outright rejection. The self-reference to the same arXiv ID in the reference list and the placeholder citations in the supplement also need editorial attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this is a competent incremental paper, not a breakthrough. The new piece is a local relation weighting (R1 plus R2) injected into a progressive feature extractor, paired with a proxy-correction module in latent space. The ablations show the local module does most of the work, and the supplement's R1/R2 comparison on three datasets is genuinely informative. The gains over AdaPoinTr are small but consistent on PCN, ShapeNet-55/34, and the projected variants. For someone working on point cloud completion, this is a useful new baseline.\n\nThe soft spots are real but mostly fixable. The KITTI fidelity number in Table 4 is the one that worries me: 0.001 (times 1e-3) against AdaPoinTr's 0.237 is a two-hundred-fold improvement, which on noisy real LiDAR is not credible under the same protocol. It reads like a typo or a different metric, and until it is explained I would not cite the KITTI claim. The missing hyperparameters M and lambda also matter, because both enter the method and the loss, and no sensitivity analysis is given. The code release could settle this, but the paper should state them. Table 1's PointAttN row is incomplete (dashes for three categories) yet still reports an average, which is sloppy. There are no error bars anywhere, so the small CD differences should be treated with some caution.\n\nNone of this is a load-bearing flaw in the architecture. The method is coherent, the ablation supports the design choices, and the benchmark comparisons follow standard protocols on the synthetic datasets. The citation pattern is normal; citing their own arXiv preprint in a supplementary reference is fine.\n\nMy recommendation: send it to peer review, but make the reviewers check the KITTI number and ask for the missing hyperparameters. If the KITTI number cannot be reproduced, the paper loses one of its four claimed datasets but the core result on PCN and ShapeNet probably still stands. I wouldn't cite it myself in the next year, but I'd bring it to a reading group discussion on benchmark hygiene in 3D completion.","headline":"Competent incremental point-cloud-completion paper with a plausible synthetic-benchmark story, undermined by an implausible KITTI fidelity number and two unreported hyperparameters.","tokens_in":16216,"tokens_out":2336,"would_cite":false,"duration_ms":25647,"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":"PointCFormer improves point cloud completion by weighting each neighbor's contribution with spatial and feature relation metrics.","keywords":["point cloud completion","transformer","local geometric relationship perception","k-nearest neighbor weighting","progressive feature extraction","point proxy correction","Chamfer distance","3D shape reconstruction"],"falsifier":"Run the released code on PCN and ShapeNet-55 while sweeping $M$ over a range such as 1, 4, 8, and 16 and $\\lambda$ over values including 0, 0.1, 1, and 10. If the reported CD-$\\ell_1$ of 6.41 on PCN and the margin over AdaPoinTr appear only at a single untested-in-public configuration, or if setting $\\lambda = 0$ or changing $M$ alters the ranking, the claim that the relation-weighted architecture drives the gain is not supported. A second check is to replace the average-edge term in $R_2$ with a fixed geometric baseline such as the local centroid and measure whether the advantage disappears.","tokens_in":15199,"feed_emoji":"🧊","tokens_out":10641,"duration_ms":92300,"temperature":0.7,"pith_summary":"PointCFormer is a transformer-based network for reconstructing a full 3D shape from an incomplete point cloud. The paper's central claim is that feature extraction improves when each neighbor's contribution to a target point is explicitly weighted by how related the two points are, measured both by spatial position and by learned feature vectors. To avoid losing global shape information during down-sampling, the network alternates global self-attention queries with a scale-tailored local feature extractor, gradually moving from dense coarse features to sparse refined ones. A correction module then re-injects dense input information into the latent point proxies. On PCN, ShapeNet-55, ShapeNet-34/Unseen-21, and KITTI, the reported Chamfer distances and F-Score are the best among the compared methods, with the largest ablation gain coming from the relation-based local perception.","feed_headline":"Per-neighbor weighting sets new best in point cloud completion","feed_subtitle":"Weighting each neighbor by spatial and feature relations beats leading methods on four 3D benchmarks.","key_machinery":"The central mechanism is the relation-based local geometric relationship perception module, which turns a standard k-NN neighborhood aggregation into a weighted aggregation. Its two metrics are $R_1(Q,V_i) = (|x_Q-x_{V_i}|, |y_Q-y_{V_i}|, |z_Q-z_{V_i}|)$, the component-wise Manhattan displacement in 3D, and $R_2(Q,V_i) = \\left|\\frac{1}{M}\\sum_{j=1}^M \\mathrm{vec}(Q,V_j) - \\mathrm{vec}(Q,V_i)\\right|$, the deviation of a neighbor's directed feature edge from the average directed edge of a small subset of neighbors. After concatenation, an MLP maps the pair to a per-neighbor contribution weight used to modulate the EdgeConv features before max-pooling. This weighted local perception is embedded in a progressive feature extractor that alternates global self-attention with scale-tailored local feature extraction, and in the point proxy correction module that pulls the latent proxies back toward the dense input distribution.","core_discovery":"The discovery the paper argues for is that k-nearest-neighbor pooling is a bottleneck in point cloud completion and can be replaced by a learned weighting scheme. For each target point, the method computes a spatial relation metric $R_1$ (Manhattan distances to each neighbor) and a feature-space relation metric $R_2$ (the absolute difference between a neighbor's directed edge and the average directed edge of a small subset of neighbors), concatenates them, and uses an MLP to produce a contribution weight per neighbor before max-pooling. This relation-based local geometric relationship perception is inserted into a progressive feature extractor that alternates self-attention and down-sampling, and into a correction module that refines the latent point proxies. The paper reports that these components together bring average Chamfer distance on PCN to 6.41 (CD-$\\ell_1$, multiplied by $10^3$) with F-Score@1\\% of 0.855, and ShapeNet-55 CD-$\\ell_2$ to 0.73, both better than the compared methods. The ablation attributes the largest part of the gain to the relation-based local perception module rather than to the global attention or correction modules.","pith_inferences":["I would read the ablation as evidence that the relation weighting is the most transferable piece: the same module produces gains in both the extractor and the correction module, so it could be plugged into other completion backbones without redesigning the whole extractor.","Because the paper does not report the value of the subset size $M$ in $R_2$ or the weight $\\lambda$ on the denoising loss, a reader should not assume the reported gains are invariant to those settings until a sensitivity sweep is published.","Since $R_2$ subtracts an average edge direction, it behaves like a local residual estimate; a natural testable variant would replace the average with a local PCA tangent direction and compare completion quality.","The two-phase progressive extractor suggests a general encoder design for downstream tasks such as segmentation or detection, where the same tension between global context and local detail appears."],"forward_implications":["If the central claim holds, k-NN-based feature extractors in other point cloud tasks can be improved by replacing unweighted pooling with relation-weighted aggregation, since the learned weighting is not specific to completion.","The reported benchmark results imply that one relation-aware extractor can simultaneously improve global shape fidelity, measured by Chamfer distance, and local detail, measured by F-Score, against AdaPoinTr, 3DMambaComplete, and PointAttN on PCN, ShapeNet-55, and ShapeNet-34.","The progressive down-sampling schedule lets dense input information be retained early and refined gradually, which the paper claims avoids both the global-structure loss of aggressive down-sampling and the high cost of a purely attention-based extractor.","The correction module implies that latent point proxies are more accurate when dense input points are explicitly re-introduced in a low-dimensional space rather than relying only on cross-attention with the encoder output."],"supporting_citations":[{"why":"Supplies the geometry-aware transformer encoder-decoder backbone and the ShapeNet-55 benchmark protocol that PointCFormer extends.","marker":"Yu et al. 2021"},{"why":"Supplies the generator, decoder scheme, loss setup, and the strongest comparison baseline; the ablation baseline is built from its partial framework.","marker":"Yu et al. 2023"},{"why":"Supplies the EdgeConv local feature aggregation that the relation-based perception reweights and that the initial coarse extractor uses.","marker":"Wang et al. 2019"},{"why":"Provides the PCN dataset and evaluation protocol used for the main benchmark and for fine-tuning on KITTI.","marker":"Yuan et al. 2018"},{"why":"Provides the recent attention-based completion baseline PointAttN, compared against in the experiments and criticized for its computational cost.","marker":"Wang et al. 2024"},{"why":"Provides the Mamba-based 3DMambaComplete baseline that PointCFormer reports beating on average Chamfer distance and F-Score on PCN.","marker":"Li, Yang, and Fei 2024"}],"fun_headline_variants":["Relation-based neighbor weights refine point cloud completion","Weighting neighbors by relations advances 3D shape completion","PointCFormer: relation-driven features top completion benchmarks","Neighbor relation metrics lift point cloud completion accuracy","Progressive extraction with relation weighting tops 3D completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported gains come from the architecture itself rather than from the choice of two unreported hyperparameters: the subset size $M$ used in the feature-relation metric $R_2$ and the weight $\\lambda$ on the denoising auxiliary loss. If those values were selected after seeing test results, the benchmark improvements could be artifacts of model selection rather than of the proposed modules.","fun_headline_variants_meta":{"raw":{"variants":["Relation-based neighbor weights refine point cloud completion","Weighting neighbors by relations advances 3D shape completion","PointCFormer: relation-driven features top completion benchmarks","Neighbor relation metrics lift point cloud completion accuracy","Progressive extraction with relation weighting tops 3D completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001463,"raw_usage":{"total_tokens":5952,"prompt_tokens":1076,"completion_tokens":4876,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":4801}},"tokens_in":692,"tokens_out":4876,"duration_ms":36264,"temperature":1.0,"reasoning_tokens":4801,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:50:57.452729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on PCN and ShapeNet-55 while sweeping $M$ over a range such as 1, 4, 8, and 16 and $\\lambda$ over values including 0, 0.1, 1, and 10. If the reported CD-$\\ell_1$ of 6.41 on PCN and the margin over AdaPoinTr appear only at a single untested-in-public configuration, or if setting $\\lambda = 0$ or changing $M$ alters the ranking, the claim that the relation-weighted architecture drives the gain is not supported. A second check is to replace the average-edge term in $R_2$ with a fixed geometric baseline such as the local centroid and measure whether the advantage disappears.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PCN dataset and evaluation protocol used for the main benchmark and for fine-tuning on KITTI."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the recent attention-based completion baseline PointAttN, compared against in the experiments and criticized for its computational cost."}],"review_version":1}