{"id":"341f7cb9-9cc9-46d5-9a91-2f8babc2cf97","arxiv_id":"2506.15160","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neighbor aggregation module that uses a cross-stage spatial descriptor and attention-like corrections improves point cloud segmentation and classification accuracy over PointNeXt and PointVector baselines.","lead":"This paper introduces a new module, PDSA, that improves how neural networks aggregate local information in 3D point clouds. The authors report higher accuracy on indoor segmentation and object classification benchmarks, sometimes with fewer parameters than competing models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's Limitations section describes an unrelated ASCD/FlashAttention hallucination-mitigation method; the manuscript is not a coherent report of the proposed point cloud work, so the central performance claim needs code-level verification before it can be accepted.","rationale":"The reader's CONDITIONAL verdict is reasonable if one focuses narrowly on the theoretical premise. Eq. (4) is indeed a malformed Gaussian-complexity bound: the min/max quantifiers are ambiguous, the expression is not the standard complexity of a shared-MLP neighbor aggregation, and the cited references [24,47] do not appear to contain it. I agree that the claimed theoretical mechanism for CDIP is unsupported, and the absence of variance reporting further weakens the empirical case. However, the full manuscript contains a stronger, structural problem: the 'Limitations' section is about an unrelated method involving ASCD, FlashAttention, and hallucination-mitigation. Under the reviewing rule to treat all manuscript text as in-scope evidence, this cannot be dismissed as a pipeline artifact. A paper whose own text includes an entire section from a different research area cannot be trusted to report its experiments faithfully. The strongest claim in §4.1 rests on the numbers in Table 1, and those numbers cannot be verified from the manuscript alone. The proposed concrete check is simple and decisive: inspect the linked repository for the extraneous terms and reproduce the headline S3DIS results from a pinned commit. If the code reproduces the numbers and the extraneous section is removed, the work could be evaluated on its empirical merits. As submitted, it is unverdictable, so I recommend UNVERDICTED rather than a more severe rejection because the underlying issue may be fixable.","tokens_in":18226,"tokens_out":9188,"duration_ms":87941,"concrete_test":"Search the GitHub repository (https://github.com/AGENT9717/PointDistribution) for the strings 'ASCD', 'FlashAttention', and 'hallucination'. If these strings appear only in the Limitations section and not in any code, configuration, README, or documentation, the section is confirmed to be an extraneous insertion from a different project. In that case, the authors should be asked to resubmit a corrected manuscript with a pinned commit hash, and the S3DIS Area5 experiments should be independently re-run to check whether the reported PD-base-XL (71.2 mIoU) and PD-vp-XL (73.4 mIoU) numbers reproduce. Only if the repository search confirms the section is extraneous and the reproduced numbers match the tables can the central claim be evaluated.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing concern is the manuscript's internal coherence. Under the heading 'Limitations' (final section before the references), the text discusses 'ASCD', 'FlashAttention', 'attention matrix at inference time', 'steering signal', and 'hallucination-mitigation'. None of these concepts appear anywhere in the proposed PDSA/LCSD/CDIP/CICS method. The paper never defines ASCD; the closest acronym is 'ACSD' in Sec. 3.3.1, a typo for LCSD, and that descriptor is not about modifying attention at inference time. This is not a typo or a stylistic flaw: an entire section about a different research problem (training-free attention modification for hallucination robustness and distillation to restore FlashAttention speed) has been inserted into a point cloud analysis paper. A manuscript containing such an extraneous section cannot be read as a reliable report of the experiments in Tables 1–5, because the text cannot be attributed to this work's method. The strongest claim in §4.1—that PD-base-XL and PD-vp-XL improve over baselines—depends entirely on those tables. Until the text is corrected and the code is verified, the central claim is not assessable. This concern is more load-bearing than the malformed Eq. (4) bound identified by the reader, because it affects the credibility of every reported number, not just the theoretical motivation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Point-Distribution Set Abstraction (PDSA), a drop-in replacement for the Set Abstraction module in PointNet++-style architectures for point cloud analysis. The method introduces a lightweight cross-stage structure descriptor (LCSD) based on octant distributions with distance weights, and two correction operations applied during neighbor aggregation: CDIP, which re-weights neighbor features to suppress irrelevant points, and CICS, which applies a global self-attention mechanism over descriptors, approximated by a key-point selection scheme, to improve class separability. The authors build two network families, PD-base and PD-vp, on top of the OpenPoint/PointNeXt and PointVector baselines, and evaluate on S3DIS semantic segmentation, ScanObjectNN and ModelNet40 classification, along with ablations and visualizations. The main empirical claim is that PD-base-XL outperforms PointNet++(OP) by 7.6% mIoU and 2.7% OA, and PointNeXt-XL by 0.7% mIoU and 0.4% OA, at roughly 70% fewer parameters, and that PD-vp-XL outperforms PointVector-XL by 0.8% mIoU and 1.1% OA.","tokens_in":18525,"tokens_out":10239,"duration_ms":100751,"significance":"If the empirical results are reproducible, the paper makes a useful and falsifiable contribution: a parameter-efficient neighbor-aggregation correction that is shown to improve several modern point cloud baselines, with a decomposed ablation study isolating CDIP, the distance weight, and CICS. The central idea of correcting the high-dimensional feature distribution via a lightweight structural descriptor, rather than adding high-dimensional geometric encodings, is plausible and worth testing, and the paper includes an explicit code/weights link, which is a strength. However, the significance of the contribution is currently limited by an unsupported theoretical premise in Eq. (4), an internally incoherent final section, absent variance reporting on small performance margins, and an underspecified key-point mechanism. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The final section, titled 'Limitations', describes an 'ASCD' method that dynamically modifies the attention matrix at inference time, its incompatibility with FlashAttention, hallucination-mitigation behavior, and distillation of a steering signal into native attention via KL divergence. None of these concepts appears in the proposed PDSA/LCSD/CDIP/CICS pipeline; the only occurrence of 'ACSD' is in Sec. 3.3.1, where it is evidently a typo for LCSD. This is not a cosmetic issue: the section reports limitations of a different research artifact, so the manuscript's text cannot be attributed to the point cloud method whose experiments appear in Tables 1-6. I ask the authors to replace this section with a genuine discussion of PDSA's limitations and to provide the released code and weights so that the reported numbers can be independently checked.","section":"Limitations (final section before References)"},{"comment":"Eq. (4) is presented as an upper bound on the Gaussian complexity of the neighbor aggregation network, attributed to references [24,47]. As written, the expression is not a valid bound: the min/max quantifiers have no declared optimization variable, the sum over j is indexed in the same expression that quantifies over p_j, the quantity depends only on input and feature coordinates rather than on the hypothesis class of the shared MLP, and neither [24] nor [47] derives this bound. Since the CDIP correction is motivated by the claim that reducing this quantity improves generalization, Eq. (4) is load-bearing. Please replace it with a correct and citable complexity bound for the 1x1-convolution aggregation used here, or explicitly reframe CDIP as a heuristic denoising operation, so that the generalization argument is not based on an invalid formula.","section":"Sec. 3.2.1, Eq. (4)"},{"comment":"The text reports 'improvements of 19.1% OA, 22.1% mAcc and 3.9% OA, 4.2% mAcc on ScanObjectNN and ModelNet40 datasets respectively' compared with PointNet++. Table 2 gives PointNet++ 77.9% OA / 75.4% mAcc on ScanObjectNN and 91.9% OA on ModelNet40, so the actual deltas for PD-S are +9.4 OA, +10.1 mAcc, and +1.2 OA; the 3.9/4.2 values are the deltas with respect to PointNet, not PointNet++. Please correct the text or the table and verify that every numeric comparison in the paper is consistent with the corresponding table.","section":"Sec. 4.2 and Table 2"},{"comment":"All experiments are reported as single runs without standard deviations or the number of seeds. Several headline claims are small margins that in this field are comparable to seed-level noise: PD-base-XL over PointNeXt-XL by +0.7 mIoU / +0.4 OA, PD-vp-XL over PointVector-XL by +0.8 mIoU / +1.1 OA, and the ablation step from CDIP+D_w to CDIP+D_w+CICS by +1.4 mIoU. Please report mean and standard deviation over at least three seeds for the main tables and ablations, and release the exact training and evaluation scripts so that the claimed improvements can be verified.","section":"Tables 1-6 (experimental protocol)"},{"comment":"The key-point mechanism is an essential part of CICS's efficiency story, but the paper never specifies the selection algorithm: it refers to 'statistical aggregates of attention weights' and 'ranking-based thresholding' without defining the statistic, the threshold, or the resulting key-point ratio. Moreover, no runtime, memory, or FLOPs comparison is presented anywhere in the experiments; the paper reports only parameter counts. The claim that key-point selection 'optimizes the computational overhead' is therefore not empirically supported. Please give a precise algorithm and complexity analysis, and report measured training/inference time and memory for PD-base-XL with and without the key-point mechanism.","section":"Sec. 3.3.3 and Sec. 4"}],"minor_comments":[{"comment":"The descriptor is introduced as 'ACSD' but is referred to as LCSD everywhere else; please use a single acronym consistently throughout the paper.","section":"Sec. 3.3.1"},{"comment":"Table 5 is difficult to read: the check-mark columns are not aligned in the typeset text, and the text says B3 and B4 gain 0.6% and 0.9% from Dw, but those deltas do not follow from the displayed B1/B2 values. Please redraw the table and reconcile the stated deltas with the numbers shown.","section":"Table 5"},{"comment":"Please fix the typo 'PoinNet++' in the text, the citation 'PointNet++ [[35]]', and the misspelled dataset name 'ScanObjecNN'.","section":"Sec. 4.2"},{"comment":"Figure 5 is referenced in the text ('Figure 5: Illustration of the network architecture') but is not present in the submitted version; please ensure the PDF contains all figures.","section":"Figure 5"},{"comment":"The notation Correct(·) is used in Eqs. (5)-(7) before the arguments r_v and r_g are defined in Sec. 3.3; consider reordering or defining these symbols earlier to improve readability.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The extraneous 'Limitations' section, the malformed Eq. (4), and the inconsistent Table 2 deltas collectively suggest that the manuscript text may have been assembled from more than one project. I would ask the editor to require a corrected manuscript and to verify the GitHub repository before further review. The paper's empirical direction is interesting, but the current submission is not a coherent report of the proposed point cloud method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2506.15160. First, the core module is a genuine, reasonably novel addition to the PointNet++ aggregation toolbox: a lightweight cross-stage octant distribution descriptor (LCSD) with a distance-aware weight, a CDIP operation that uses a cross-attention-style weight to correct the neighbor feature matrix, and a CICS operation that applies global self-attention only on selected key points. The ablations are coherent and the directions are consistent: CDIP alone adds 5.1 mIoU on S3DIS Area5, the distance weight adds 1.1, CICS adds 1.4. The parameter-efficiency comparison against PointNeXt-XL is not apples-to-apples, but it is honestly reported (12.5M vs 41.6M, so ~70% reduction).\n\nSecond, more worrying: the final 'Limitations' section describes a completely different research problem. It talks about ASCD modifying the attention matrix at inference time for hallucination mitigation and FlashAttention incompatibility. None of that appears anywhere in the point cloud method. That is not a small typo; someone pasted a section from an unrelated paper into this manuscript. The stress-test is right: this undermines the credibility of the experimental tables, because we cannot be sure the text was assembled with care.\n\nThe reader's main technical concern, Eq. (4), is also legitimate but softer. The equation is presented as an established Gaussian-complexity bound, but it is malformed and the cited references do not support it. That weakens the stated motivation for CDIP, but the ablation evidence for CDIP stands on its own, so I treat it as a motivation gap, not a fatal flaw.\n\nOther soft spots: all results are single-run with no variance, the validation fold is used for model selection, and there are scattered typos ('ACSD', 'CISC', 'pooing'). All fixable.\n\nMy take: the architecture is worth looking at, and the empirical pattern is plausible. But the manuscript as submitted is not a coherent report, and the numbers cannot be trusted until the authors provide a corrected version and working code with a commit hash. I would send it to peer review intending to request both, and let the reviewers verify. If the code reproduces the tables, this becomes a modest but solid contribution. If it doesn't, the whole thing falls apart.","headline":"A plausible new PointNet++ aggregation module with coherent ablations, but an unrelated 'Limitations' section pasted into the manuscript means the numbers need code verification before anyone should trust them.","tokens_in":19074,"tokens_out":3512,"would_cite":false,"duration_ms":35572,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neighbor-aggregation correction based on cross-stage structure correlation improves point cloud segmentation and classification accuracy.","keywords":["point cloud analysis","neighbor aggregation","set abstraction","feature distribution correction","cross-stage structure descriptor","semantic segmentation","point cloud classification","self-attention"],"falsifier":"Run a capacity-matched ablation: replace CDIP with an additive correction of equal parameter and flop budget that does not reduce neighbor-feature variance, keeping LCSD and CICS fixed, and measure S3DIS mIoU. If the variance-reducing version does not beat the capacity-matched alternative, the paper's stated mechanism is not what drives the gain. A complementary check is to compute a conventional Rademacher or Gaussian complexity bound for a randomly initialized shared-MLP aggregation block on a fixed neighborhood and test whether the min-max quantity in Eq. (4) predicts observed generalization across neighborhoods with different feature variance.","tokens_in":17962,"feed_emoji":"🧊","tokens_out":10761,"duration_ms":97302,"temperature":0.7,"pith_summary":"The paper argues that the standard recipe for aggregating local point-cloud neighborhoods—encoding the relative three-dimensional coordinates of sampled neighbors and pooling them—is hampered by two kinds of noise: irrelevant points that enter the sampled neighborhood, and a mismatch between low-dimensional geometry and high-dimensional semantic features. To fix this, it proposes the Point-Distribution Set Abstraction (PDSA) module, which first builds a lightweight cross-stage descriptor (LCSD) of each point's spatial distribution and then uses that descriptor to correct the neighbor-feature distribution during aggregation. One correction (CDIP) suppresses irrelevant points by shrinking the variance of the neighbor feature matrix; the other (CICS) models long-range context to keep different classes separated. On S3DIS semantic segmentation, the PD-base-XL model, built entirely from PDSA modules, beats the strengthened PointNet++ baseline by 7.6 percent mIoU and 2.7 percent OA, and beats PointNeXt-XL by 0.7 percent mIoU with about 70 percent fewer parameters; on ScanObjectNN and ModelNet40 classification, the PD-S model reaches 87.3 and 93.1 percent OA. The package matters because it tries to improve local aggregation by correcting distributions in feature space rather than by adding expensive high-dimensional geometric encodings.","feed_headline":"New point-cloud module beats PointNet++ by 7.6% mIoU","feed_subtitle":"It corrects feature distributions during neighbor pooling, using about 70% fewer parameters than PointNeXt-XL.","key_machinery":"The load-bearing mechanism is the Point-Distribution Set Abstraction (PDSA) module, built on a Lightweight Cross-stage Structure Descriptor (LCSD). The LCSD encodes, for each point, the distribution of its neighbors across eight spatial octants, with distance-aware relative weights and a dimension-reducing linear layer that lets the descriptor be propagated across stages. CDIP (correction for denoising irrelevant points) uses the difference between the next-stage neighbor descriptor and each point's descriptor, in a cross-attention-inspired formulation, to compute an attention weight and a structural code added into the neighbor feature matrix, thus reducing its per-channel variance. CICS (correction for improving class separability) applies global self-attention to the descriptor to pull in long-range context; to keep cost down, a key-point selection mechanism ranks points by aggregated attention weight and computes the global correction only at those points, copying it to the rest of the neighbor. The paper's Eq. (7) summarizes the combined correction: pooled neighbor features are corrected both inside the neighbor matrix and at the pooled output.","core_discovery":"On the paper's own terms, the central claim is that neighbor aggregation in point cloud networks can be improved by correcting the distribution of aggregated features in high-dimensional space using a cross-stage structure descriptor, rather than by refining low-dimensional geometric descriptions. The PDSA module operationalizes this: a lightweight descriptor aggregates octant-wise distribution features of each neighborhood with distance-aware weights across network stages; the CDIP operation compares each neighbor's descriptor with the next-stage neighbor descriptor to generate attention weights and structural codes that reduce the variance of the neighbor feature matrix; and the CICS operation applies a global self-attention over the descriptor, restricted to automatically selected key points, to sharpen class separability. The paper reports that a network composed entirely of PDSA modules (PD-base-XL) reaches 71.2 percent mIoU and 91.0 percent OA on S3DIS Area 5 with 12.5 million parameters, and that adding PDSA to the PointVector baseline (PD-vp-XL) reaches 73.4 percent mIoU and 91.8 percent OA. These numbers are presented as evidence that distribution correction during aggregation generalizes across baselines and tasks.","pith_inferences":["Editorial inference: because the LCSD is an octant histogram with distance weights, it is aligned to the global coordinate axes; on rotation-augmented or orientation-varied inputs the descriptor itself rotates, so a testable extension is to make the descriptor rotation-invariant by canonicalizing the octants with a local frame and see whether the reported gains persist.","Editorial inference: the key-point mechanism selects points by aggregated attention weights, so in highly class-imbalanced scenes the selected key points may under-represent rare classes; a class-balanced key-point sampling variant would test whether the CICS correction has a class-bias failure mode.","Editorial inference: the theoretical motivation for CDIP is not tightly established, since the cited Gaussian-complexity bound is not reproduced for shared-MLP aggregation; the empirical gains could in principle come from the added capacity of the attention-style correction rather than from variance reduction, and a capacity-matched ablation would separate these.","Editorial note on the manuscript text: the appended 'Limitations' paragraph concerns an unrelated ASCD/FlashAttention/hallucination-mitigation mechanism, not the PDSA point-cloud method, so it should not be read as a stated limitation of the claims here."],"forward_implications":["Replacing standard Set Abstraction modules with PDSA improves S3DIS semantic segmentation: PD-base-XL reports 71.2 percent mIoU and 91.0 percent OA with 12.5M parameters, versus 63.6 percent mIoU for the PointNet++(OP) baseline, and PD-vp-XL reports 73.4 percent mIoU and 91.8 percent OA, above the PointVector-XL baseline.","The ablation attributes the gain mainly to CDIP, which adds 5.1 percent mIoU over the PointNet++ baseline; the distance-aware weight in the LCSD adds 1.1 percent, and CICS adds 1.4 percent.","The module generalizes to classification: PD-S reaches 87.3 percent OA and 85.5 percent mAcc on ScanObjectNN, and 93.1 percent OA and 90.4 percent mAcc on ModelNet40 with 2.5M parameters.","Using PDSA in the local-aggregation block of an existing baseline (PointVector) improves that baseline by 0.8 percent mIoU and 1.1 percent OA, suggesting the correction is modular rather than tied to one architecture.","Because PDSA compares each point with the overall neighbor distribution rather than pairwise point similarities, it is designed to handle the irregular neighborhoods produced by K-NN or ball-query grouping."],"supporting_citations":[{"why":"Defines the Point Set Abstraction module and neighbor-aggregation pipeline that the paper's PDSA replaces and improves.","marker":"[34]"},{"why":"Supplies the strengthened PointNet++(OP) baseline, the PointNeXt-XL comparison, and the training recipe used across experiments.","marker":"[35]"},{"why":"Provides the PointVector baseline and VPSA local aggregation block that the PD-vp variant augments with PDSA.","marker":"[9]"},{"why":"Prior neighbor-denoising method that motivates CDIP, and one of the two cited sources for the Gaussian-complexity bound in Eq. (4).","marker":"[47]"},{"why":"Cited together with [47] as the source of the Eq. (4) bound that motivates minimizing neighbor-feature variance via CDIP.","marker":"[24]"},{"why":"Contributes the octant-centroid cross-stage encoding that the LCSD descriptor adapts with distance-aware weights.","marker":"[59]"},{"why":"Establishes explicit cross-stage structure modeling and the non-Euclidean manifold argument that motivates the feature-separability correction.","marker":"[39]"},{"why":"Provides the S3DIS benchmark used for the main semantic segmentation results and ablations.","marker":"[1]"},{"why":"Provides the ScanObjectNN classification benchmark on which PD-S is evaluated.","marker":"[41]"},{"why":"Provides the ModelNet40 classification benchmark on which PD-S is evaluated.","marker":"[51]"}],"fun_headline_variants":["Point cloud module lifts mIoU by 7.6% with fewer parameters","PDSA: Correct neighbor features to boost point cloud segmentation","Distribution correction improves point cloud segmentation accuracy","New module enhances point cloud analysis with fewer parameters","Cross-stage structure correlation boosts point cloud segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reducing the Euclidean spread of neighbor-point features inside each aggregated neighborhood is what improves generalization (Eq. 4); if that bound is not valid for shared-MLP aggregation, the theoretical motivation for the CDIP correction is unsupported, even though the module could still work empirically.","fun_headline_variants_meta":{"raw":{"variants":["Point cloud module lifts mIoU by 7.6% with fewer parameters","PDSA: Correct neighbor features to boost point cloud segmentation","Distribution correction improves point cloud segmentation accuracy","New module enhances point cloud analysis with fewer parameters","Cross-stage structure correlation boosts point cloud segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3394,"prompt_tokens":1014,"completion_tokens":2380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":2303}},"tokens_in":630,"tokens_out":2380,"duration_ms":16721,"temperature":1.0,"reasoning_tokens":2303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:42:45.880625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a capacity-matched ablation: replace CDIP with an additive correction of equal parameter and flop budget that does not reduce neighbor-feature variance, keeping LCSD and CICS fixed, and measure S3DIS mIoU. If the variance-reducing version does not beat the capacity-matched alternative, the paper's stated mechanism is not what drives the gain. A complementary check is to compute a conventional Rademacher or Gaussian complexity bound for a randomly initialized shared-MLP aggregation block on a fixed neighborhood and test whether the min-max quantity in Eq. (4) predicts observed generalization across neighborhoods with different feature variance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Point Set Abstraction module and neighbor-aggregation pipeline that the paper's PDSA replaces and improves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the strengthened PointNet++(OP) baseline, the PointNeXt-XL comparison, and the training recipe used across experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PointVector baseline and VPSA local aggregation block that the PD-vp variant augments with PDSA."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior neighbor-denoising method that motivates CDIP, and one of the two cited sources for the Gaussian-complexity bound in Eq. (4)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited together with [47] as the source of the Eq. (4) bound that motivates minimizing neighbor-feature variance via CDIP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the octant-centroid cross-stage encoding that the LCSD descriptor adapts with distance-aware weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes explicit cross-stage structure modeling and the non-Euclidean manifold argument that motivates the feature-separability correction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the S3DIS benchmark used for the main semantic segmentation results and ablations."},{"cited_title":"In2024 International Confer- ence on 3D Vision (3DV), pages 179–189","cited_arxiv_id":null,"evidence_quote":"Provides the ScanObjectNN classification benchmark on which PD-S is evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ModelNet40 classification benchmark on which PD-S is evaluated."}],"review_version":2}