{"id":"5632ecb9-4565-4ece-8033-ef65ee567369","arxiv_id":"2411.15066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A point cloud completion network that uses boundary points (the interface) as a structural prior and predicts coarse shapes as displacements of those boundary points achieves lower Chamfer distance than prior methods on ShapeNet, PCN, and KITTI.","lead":"SPAC-Net completes missing regions of point clouds by first locating boundary points between the observed and missing surface, then predicting how those boundary points move into the missing region. It outperforms several prior networks on three benchmarks while using fewer FLOPs, but its best results rely on knowing the camera viewpoint used to create the partial scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ShapeNet SOTA claim rests on interface localization using the known occlusion viewpoint, which baselines lack and the method's own PCN/KITTI pipeline avoids; the paper's conclusion admits automatic localization underperforms, so the headline comparison is not yet clean.","rationale":"The paper proposes a genuinely interesting structural prior: instead of directly regressing missing points, it localizes the boundary between observed and missing regions and learns displacements from that boundary. The architecture is coherent, the ablations (Tables 6, 7, 8) support the module choices, and the PCN and KITTI results use viewpoint-free localization, which is a meaningful positive sign. The authors also provide code, which aids verification. However, the single most load-bearing assumption is that the ShapeNet benchmark comparison is fair. The interface is localized using the exact occlusion viewpoint that generated each partial scan, a quantity unavailable to the competing methods and unavailable in the authors' own real-world pipeline. This is not a minor implementation detail: the interface is the driving signal of the method, and Sec. 5 explicitly concedes that automatic localization performed worse than the MAD detector. Without a ShapeNet ablation using viewpoint-free localization, the reported gains over SeedFormer and PoinTr could be entirely attributable to the privileged information rather than to the architectural contribution. The reader's conditional verdict already identifies this concern, and I agree. I recommend keeping the conditional verdict pending the concrete test: if the viewpoint-free ShapeNet result preserves the gains, the paper should be accepted; if not, the central claim needs to be reframed around PCN/KITTI or the known-viewpoint setting must be disclosed as an additional supervision signal available only to SPAC-Net. Secondary concerns, such as unspecified NT and r values and test-set tuning of delta, are reproducibility issues that should also be fixed but are less central than the benchmark fairness question.","tokens_in":15846,"tokens_out":3155,"duration_ms":32128,"concrete_test":"Re-run the ShapeNet-55/34 experiments with the viewpoint-free edge detector of Sec. 3.2 (Fig. 4b) in place of the known-occlusion-point selection (Sec. 3.2, case 1, Fig. 4a), for both SPAC-Net and PoinTr+interface, retraining with the same hyperparameters; report CD-S/M/H, CD-Avg, and F1. If the advantage over SeedFormer (CD-Avg 0.88 vs 0.92) does not survive this unbiased interface localization, the headline claim is not supported by a fair comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SPAC-Net outperforms state-of-the-art methods is anchored in the ShapeNet-55/34 tables (Tables 1-2). Those benchmarks are generated by occluding points farthest from a known viewpoint (Sec. 4.1), and the MAD module localizes the interface by selecting the NT points nearest to that known occlusion point (Sec. 3.2, case 1, Fig. 4a). This is privileged information: FoldingNet, PCN, PoinTr, and SeedFormer do not receive it, and SPAC-Net's own PCN and KITTI evaluations use the viewpoint-free edge detector of Sec. 3.2 (case 2, Fig. 4b). The comparison is therefore not apples-to-apples. The paper's conclusion (Sec. 5) states that automatic end-to-end interface localization did not perform as well as MAD, which is consistent with the known viewpoint providing the measurable advantage. If the ShapeNet gains vanish when interface localization is viewpoint-free, the 'outperforms SOTA' claim is overstated for that benchmark; the PCN/KITTI results may still stand but do not validate the ShapeNet comparison. Secondary reproducibility gaps (NT and r are not given numeric values; delta in Table 7 is tuned directly on the PCN test set) reinforce the need for a clean, fully specified evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPAC-Net, a point cloud completion framework built around a new structural prior called the 'interface', defined as the boundary between the observed partial scan and the missing region. A Marginal Detector (MAD) module localizes this interface, either from a known occlusion viewpoint (ShapeNet-55/34) or from a viewpoint-free edge detector (PCN, KITTI). The network then predicts a coarse shape by learning displacements from interface points to their corresponding positions in the missing part, refines the coarse shape with a Structure Supplement (SSP) module, and finally applies a FoldingNet-based upsampler. The paper reports state-of-the-art results on ShapeNet-55/34, PCN, and KITTI, with ablations supporting the interface definition and the number of SSP modules.","tokens_in":16118,"tokens_out":5474,"duration_ms":47938,"significance":"If the results hold, the idea of explicitly conditioning coarse-shape generation on boundary points is a useful and generally applicable contribution to point cloud completion. The PCN and KITTI evaluations use a viewpoint-free interface detector, and the ablations (Tables 6 and 8) support the proposed module choices. The code is publicly available, which aids reproducibility. However, the headline ShapeNet-55/34 comparison is currently compromised by the use of privileged viewpoint information during interface localization, and two core hyperparameters are left unspecified; these issues must be resolved before the claimed state-of-the-art performance can be taken at face value.","major_comments":[{"comment":"The ShapeNet-55/34 comparison is not apples-to-apples: in these benchmarks the interface is localized by the MAD module using the known occlusion viewpoint that generated each partial scan (Sec. 3.2, case 1), while the baselines (FoldingNet, PCN, PoinTr, SeedFormer) do not receive this privileged information. The paper's own conclusion (Sec. 5) states that automatic end-to-end interface localization 'did not perform as well as the MAD method', which is consistent with the known viewpoint providing a measurable advantage. To support the claim of outperforming state-of-the-art approaches on ShapeNet-55/34, please report results on these benchmarks using the viewpoint-free edge detector of Sec. 3.2 (case 2), or explicitly label the current ShapeNet numbers as oracle-interface results and restrict the headline comparison to the PCN and KITTI benchmarks.","section":"Sec. 4.3, Tables 1-2"},{"comment":"The hyper-parameter δ of the edge detector is selected by evaluating four values on the PCN test set and reporting the one with the best CD-ℓ1 (δ=0.5). This constitutes test-set tuning: the reported 6.61 average is the best of four runs on the test split, which inflates the result relative to the fixed-δ deployment used by the comparison methods. Please select δ on a validation split (or report the results for all four δ values as a sensitivity analysis), and likewise specify the exact split used.","section":"Sec. 4.6, Table 7"},{"comment":"The number of interface points NT and the edge-detection radius r are core parameters of the MAD module but are never given numeric values. NT is used in Fig. 3 and in the feature-dimension descriptions (NP=NT, CP=256; edgeconv(N=NT, CT=3), etc.), and r is introduced in Sec. 3.2 as the radius controlling the neighborhood, but neither is specified for ShapeNet, PCN, or KITTI. Without these values the experiments cannot be reproduced; please state them explicitly.","section":"Sec. 4.2"},{"comment":"The localization rule for the occlusion-available scenario states: 'By selecting the nearest NT points in terms of distance from the occlusion point, we can obtain the interface.' Under the benchmark generation described in Sec. 4.1, the partial scan is obtained by removing the N points farthest from a viewpoint, so the interface (the boundary between the partial scan and the missing part) consists of the points in the partial scan with the largest distance from the viewpoint, not the smallest. As written, the rule selects the innermost points rather than the boundary. Please clarify whether the implementation actually selects the farthest points, or explain the geometry in which the nearest points form the boundary.","section":"Sec. 3.2, case 1"}],"minor_comments":[{"comment":"The phrase 'suffer from the loss of details inevitably' is awkward; consider rephrasing for clarity.","section":"Abstract"},{"comment":"The subscript/superscript in FPiTi appears to be a typo; it should be the relative feature FPT between partial scans and interfaces introduced in Sec. 3.2.","section":"Eq. (4)"},{"comment":"The Fidelity values of 0.000 for PoinTr and SPAC-Net are suspicious; please state whether the output always contains the input points, which makes Fidelity trivially zero.","section":"Table 4"},{"comment":"FLOPs is a count, not a rate; the parenthetical 'Floating Point Operations per Second' is a misnomer and should be corrected.","section":"Sec. 4.5"},{"comment":"The optimizer name is written as 'ADAMW'; it should be 'AdamW'.","section":"Sec. 4.2"},{"comment":"The word 'symetric' should be 'symmetric'.","section":"Sec. 3.5"}],"recommendation":"major_revision","confidential_remarks":"The privileged-information concern in Tables 1-2 is the main risk. If the authors cannot provide a fair comparison using the viewpoint-free detector on ShapeNet, the paper's central claim is overstated. The test-set tuning of δ and the missing NT/r values would likely be raised by other reviewers as well. I recommend requesting a major revision with a clear protocol for the ShapeNet comparison and full disclosure of all interface-localization hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nSPAC-Net introduces a genuinely new structural prior—the interface between observed and missing regions—and uses it to predict the coarse shape as displacements from interface points. That is not AnchorFormer in disguise; AnchorFormer's anchors are learned, while the interface is localized boundary geometry. The SSP refinement module is sensible, and the ablations in Tables 6 and 8 support the design choices. The efficiency result is real: 5.2 GFLOPs versus SeedFormer's 29.6, with fewer parameters than PoinTr. The authors also test \"PoinTr + interface\" as a control, which is a good check that the prior helps a strong baseline. The limitations section is unusually honest.\n\nThe serious problem is the ShapeNet-55/34 evaluation. Partial scans are generated by removing points farthest from a known viewpoint, and the MAD module localizes the interface by selecting the NT points nearest to that same viewpoint. That is privileged information. FoldingNet, PCN, PoinTr, and SeedFormer never receive it, and the paper's own PCN and KITTI experiments use the viewpoint-free edge detector instead. So the headline \"outperforms SOTA\" on ShapeNet is not an apples-to-apples comparison. The conclusion makes this worse by admitting that end-to-end automatic localization \"did not perform as well as the MAD method.\" That is what you would expect if the known viewpoint is providing the advantage. The PCN results (average L1-CD 6.61 vs 6.74) and KITTI MMD (0.473 vs 0.516) are still positive, but they do not validate the ShapeNet comparison.\n\nSmaller issues: NT and the edge-detection radius r are never given numeric values; delta in Table 7 is tuned directly on the PCN test set; and the KITTI \"Fidelity 0.000\" for both PoinTr and SPAC-Net is trivially achieved by concatenating the input points back, so it shouldn't count as evidence. None of these are fatal alone.\n\nMy recommendation: send to peer review. The idea is worth engaging with, the efficiency is practically useful, and the PCN/KITTI evidence suggests the method works even without the privileged viewpoint. But the authors must re-run ShapeNet with viewpoint-free interface localization, or at least present the known-viewpoint setting separately with an explicit fairness discussion, report NT and r, and stop tuning delta on the test set. As written, the \"outperforms SOTA\" claim is overstated for ShapeNet.","headline":"New interface prior with real efficiency gains, but the ShapeNet SOTA claim is inflated by privileged occlusion-viewpoint information.","tokens_in":16704,"tokens_out":4126,"would_cite":true,"duration_ms":38003,"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":"SPAC-Net claims that locating the boundary between a partial scan and its missing region, then moving those boundary points into the gap, yields state-of-the-art point cloud completion.","keywords":["point cloud completion","structural prior","interface","displacement learning","point cloud transformer","3D shape reconstruction","detail preservation"],"falsifier":"Run SPAC-Net on ShapeNet-55/34 with the MAD edge detector or an end-to-end interface predictor instead of the known occlusion viewpoint, and compare Chamfer distance and F-Score to the paper's reported numbers; if the margin over SeedFormer and PoinTr disappears, the reported gains depend on privileged viewpoint information, not on the interface prior itself.","tokens_in":15642,"feed_emoji":"🧩","tokens_out":6193,"duration_ms":51159,"temperature":0.7,"pith_summary":"The paper tries to establish that point cloud completion — inferring a full 3D shape from a partial scan — is done better when the network is guided by an explicit structural prior called the interface: the set of points where the known observation meets the missing region. Instead of generating the coarse shape from a global feature vector, SPAC-Net learns displacements that carry interface points into their corresponding positions in the missing part. A second module, the Structure Supplement (SSP), refines the coarse shape's structural detail before upsampling. On ShapeNet-55/34, PCN, and KITTI, the paper reports lower Chamfer distance and higher F-Score than previous state-of-the-art methods, including PoinTr and SeedFormer. If this is right, explicitly locating the boundary between known and unknown geometry is a stronger inductive bias for completion than learning purely from global shape encodings.","feed_headline":"Move the boundary, fill the gap: new point cloud completion prior","feed_subtitle":"SPAC-Net locates the scan's boundary, shifts those points into the hole, and beats SeedFormer and PoinTr.","key_machinery":"The load-bearing object is the interface, defined as the set of points lying on the shared boundary between the known partial scan and the missing region; the paper treats the interface as a structural prior because, in the limit of infinitesimal point spacing, the interface computed from the observed side equals the interface on the missing side. The MAD module localizes the interface by viewpoint-distance selection when the occlusion point is known and by a two-threshold edge detector when it is not, and coarse-shape generation is redefined so that each interface point is displaced by a learned MLP-transformed relative feature and added to the interface coordinates. The SSP module then iteratively injects partial-scan detail into the coarse shape via multi-head self- and cross-attention before the FoldingNet upsampler produces the final dense shape.","core_discovery":"The paper's central claim is that the boundary between the observed partial scan and the unobserved missing part — the interface — is a reliable structural prior, and that treating coarse-shape generation as a movement of interface points into the missing region preserves local detail that encoder-decoder pipelines lose through max-pooling. The authors define the interface as the intersection between the partial scan and the missing part, argue that both sides share the same boundary as the sampling radius approaches zero, and localize it with a Marginal Detector (MAD) module — using the known occlusion viewpoint on ShapeNet-55/34 and a viewpoint-free edge detector on PCN and KITTI. The predicted coarse shape is then refined by stacked Structure Supplement (SSP) modules that combine self- and cross-attention with the partial-scan features, before a FoldingNet-based upsampler produces the dense completion. The paper reports that SPAC-Net outperforms existing state-of-the-art methods on the three benchmarks, and that adding the interface to a PoinTr baseline improves that baseline as well.","pith_inferences":["An implication the paper leaves implicit is that the ShapeNet-55/34 results depend on the occlusion viewpoint being known; the paper does not report a ShapeNet ablation using the viewpoint-free edge detector, so the practical gain in real settings without known viewpoints remains untested.","The interface idea could be adapted to other geometry-recovery tasks such as point cloud denoising or inpainting, wherever a boundary between reliable and corrupted regions can be defined.","An end-to-end learned interface localizer is a natural next step; the paper notes that its automatic localization experiments underperformed MAD, leaving open whether a better-trained localizer would close the gap to the viewpoint-supervised version.","A testable extension would be to add controlled noise to the interface points on synthetic data and measure how completion accuracy degrades, quantifying how much of the gain comes from precise boundary localization."],"forward_implications":["If correct, completion networks no longer need to hallucinate missing geometry from a global feature; they can anchor prediction to observed boundary points, which should reduce detail loss in thin structures and edges.","Adding the interface as an auxiliary prior to an existing transformer baseline (PoinTr + interface) improves its Chamfer distance and F-Score, suggesting the prior is transferable rather than architecture-specific.","The two-phase design — coarse shape via displacement plus SSP refinement before upsampling — implies upsampling modules can concentrate on densification instead of also repairing structure, consistent with the paper's reported lower FLOPs.","On unseen categories in ShapeNet-34, the method reports better F-Score than SeedFormer, implying the boundary prior helps generalization to novel object classes."],"supporting_citations":[{"why":"Supplies the ShapeNet-55/34 benchmark, the PoinTr baseline, and the 'PoinTr + interface' variant that isolates the effect of the interface prior.","marker":"[44]"},{"why":"Provides the SeedFormer baseline and the encoder design using patch seeds with point transformer and set abstraction that SPAC-Net follows.","marker":"[53]"},{"why":"Supplies the PCN benchmark and its partial-scan generation protocol, used for the high-resolution completion evaluation.","marker":"[45]"},{"why":"Supplies the FoldingNet upsampler used to disperse dense points in the final stage.","marker":"[43]"},{"why":"Provides EdgeConv, used to extract the relative feature between interfaces and partial scans for displacement learning.","marker":"[38]"},{"why":"Supplies set abstraction layers used in the encoder to extract local features from partial scans.","marker":"[25]"},{"why":"Supplies the point transformer layers used in the encoder.","marker":"[50]"},{"why":"Defines the F-Score metric used to evaluate completion accuracy.","marker":"[31]"},{"why":"Supplies the KITTI dataset used for the real-world evaluation with Fidelity and MMD metrics.","marker":"[12]"}],"fun_headline_variants":["SPAC-Net uses the scan boundary as a prior to fill holes","Move boundary points to complete point clouds: SPAC-Net","Boundary as prior: SPAC-Net completes point clouds by shifting interface points","Interface prior helps SPAC-Net beat state-of-the-art for point cloud completion","Shifting the boundary: SPAC-Net's structural prior for point cloud completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main benchmark result rests on the assumption that the network knows the exact occlusion viewpoint on ShapeNet-55/34, so it can select interface points by distance to that viewpoint; competing methods do not receive this information, and the paper reports no ShapeNet ablation using its viewpoint-free localization.","fun_headline_variants_meta":{"raw":{"variants":["SPAC-Net uses the scan boundary as a prior to fill holes","Move boundary points to complete point clouds: SPAC-Net","Boundary as prior: SPAC-Net completes point clouds by shifting interface points","Interface prior helps SPAC-Net beat state-of-the-art for point cloud completion","Shifting the boundary: SPAC-Net's structural prior for point cloud completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3551,"prompt_tokens":964,"completion_tokens":2587,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2487}},"tokens_in":580,"tokens_out":2587,"duration_ms":16166,"temperature":1.0,"reasoning_tokens":2487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:35:04.738855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SPAC-Net on ShapeNet-55/34 with the MAD edge detector or an end-to-end interface predictor instead of the known occlusion viewpoint, and compare Chamfer distance and F-Score to the paper's reported numbers; if the margin over SeedFormer and PoinTr disappears, the reported gains depend on privileged viewpoint information, not on the interface prior itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ShapeNet-55/34 benchmark, the PoinTr baseline, and the 'PoinTr + interface' variant that isolates the effect of the interface prior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SeedFormer baseline and the encoder design using patch seeds with point transformer and set abstraction that SPAC-Net follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PCN benchmark and its partial-scan generation protocol, used for the high-resolution completion evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FoldingNet upsampler used to disperse dense points in the final stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides EdgeConv, used to extract the relative feature between interfaces and partial scans for displacement learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies set abstraction layers used in the encoder to extract local features from partial scans."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the point transformer layers used in the encoder."},{"cited_title":"Tatarchenko, S","cited_arxiv_id":null,"evidence_quote":"Defines the F-Score metric used to evaluate completion accuracy."},{"cited_title":"Geiger, P","cited_arxiv_id":null,"evidence_quote":"Supplies the KITTI dataset used for the real-world evaluation with Fidelity and MMD metrics."}],"review_version":1}