{"id":"79d9ed84-5f9e-4862-acea-39c7745e0032","arxiv_id":"2501.10966","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DC-PCN adds dual-level codebook quantization and an information exchange mechanism to a transformer-based point cloud completion network, reporting new state-of-the-art CD and F-Score results on PCN, ShapeNet_Part, and ShapeNet34.","lead":"A new point cloud completion network, DC-PCN, uses two codebooks to quantize features at two levels, aiming to reduce ambiguity from different samplings of the same 3D surface. It reports small but consistent improvements over prior state-of-the-art methods on four benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on per-dataset codebook-size tuning and single-run margins; without a validation protocol or variance estimates, the 0.07 CD-l1 gain over AdaPoinTr cannot be distinguished from selection noise.","rationale":"The reader identified hyperparameter selection risk as the weakest assumption, and my reading of the paper points to the same condition. The full pipeline's advantage over both its own no-codebook baseline and the strongest prior method is exactly 0.07 CD-l1 on PCN, and the ablation table reports no variance. The Limitations section explicitly says codebook sizes were tuned to optimal values, but the tuning protocol is not disclosed. If that tuning used test-set labels, or if the single-run numbers are noisy at the 0.01 scale, the SOTA claim collapses. I considered whether the more technical issue of Eq. (4) being an unbounded repulsion loss is a stronger concern, but the paper's own experiments are the primary evidence for the central claim, and that evidence is fragile for the reasons above. Therefore I do not move the verdict; conditional acceptance pending disclosure of the tuning protocol and a variance check remains appropriate. I give credit where due: the architecture is clearly described, the ablations cover each module, and the reported metrics are internally consistent across tables, which makes the remaining risk specifically about selection bias and statistical significance rather than about fabrication or mathematical invalidity.","tokens_in":13146,"tokens_out":5900,"duration_ms":63334,"concrete_test":"Re-run the PCN experiments for methods A (no codebook), E (full), and AdaPoinTr under the same training schedule with 10 random seeds, keeping the codebook size fixed at the value used for E, and report mean ± std of CD-l1 and F-Score. If the E-vs-A gap (0.07) or the E-vs-AdaPoinTr gap (0.07) is smaller than one pooled standard deviation, the reported SOTA margin is not statistically distinguishable from noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: DC-PCN beats AdaPoinTr by 0.07 CD-l1 on PCN and by smaller margins on the other benchmarks. The authors' Limitations paragraph admits that \"the sizes of codebooks in the experiments are all the optimal values obtained by changing different parameters,\" but the paper never reports the search space, the selection criterion, or whether the test split was touched during that search. This matters because every reported gain is small: the full pipeline (E) improves over the no-codebook baseline (A) by exactly 0.07 on PCN (Table 5), the same magnitude as the headline improvement over AdaPoinTr; the QIE module alone contributes 0.01 and the decoder-codebook 0.01. The ablation reports a single run with no error bars, so any run-to-run variance of the order of 0.01-0.07 would erase both the internal ablation conclusions and the SOTA margin. This is not an internal contradiction, but it is the least secure load-bearing condition: if the optimal K values were selected by test performance, or if single-run numbers vary at the reported effect sizes, the claim that dual-codebook quantization and QIE improve completion is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DC-PCN, an encoder-decoder point cloud completion network that introduces two vector-quantization codebooks: an encoder-codebook applied to shallow regional features before the transformer encoder, and a decoder-codebook applied to deep features before the shape predictor. A quantized information exchange (QIE) module deduplicates, re-targets, and merges codes from the two codebooks, and a contrastive-style codebook loss is added to Chamfer-distance supervision. The method is evaluated on PCN, ShapeNet_Part, ShapeNet34, and KITTI, reporting the best or competitive Chamfer distance and F-Score, with a claimed state-of-the-art average CD-l1 of 6.46 on PCN. Ablations on PCN (Table 5) decompose the gains attributable to the encoder codebook, decoder codebook, and QIE.","tokens_in":13360,"tokens_out":5339,"duration_ms":52666,"significance":"If the reported results are robust, the dual-codebook design is a meaningful extension of VQ-based point cloud completion: quantizing at both shallow and deep levels and exchanging information between codebooks targets the sampling-ambiguity problem, and the paper provides a clearly specified loss structure and an ablation that includes a shared-codebook control. The authors also release code, which supports reproducibility. The central weakness is empirical: all gains over the baseline and over prior state of the art are 0.01--0.07 in the primary metric, come from single runs with no variance estimates, and depend on per-dataset codebook sizes whose selection protocol is not described. The stress-test concern about small margins and codebook tuning therefore lands, and it is load-bearing for the headline claim.","major_comments":[{"comment":"The state-of-the-art claim is not supported by the evidence as reported. The full model improves over the no-codebook baseline A by 0.07 CD-l1 (6.53 to 6.46) in Table 5, which is exactly the size of the reported improvement over AdaPoinTr in Table 1 (6.53 to 6.46). The individual ablations add only 0.05 (EC), 0.01 (DC), and 0.01 (QIE), and no error bars or multiple-seed results are provided. Run-to-run variation of this magnitude would erase both the internal ablation conclusions and the headline margin. The Limitations paragraph further admits that the codebook sizes are 'the optimal values obtained by changing different parameters,' but the search space, the selection criterion, and whether the test split was used are not specified. The authors should report multiple runs with mean and standard deviation (or equivalent statistical evidence) and a validation-based hyperparameter selection protocol, or the state-of-the-art claim should be withdrawn.","section":"Limitations; Table 1; Table 5"},{"comment":"The central hyperparameters are not reported. The paper never gives the codebook size K or dimension R, the number of regions M, the deep-feature dimensions H and C, the optimizer, learning rate, batch size, training epochs, or the region/FPS parameters, so the method cannot be reproduced from the text; the only implementation information is the GPU type. Because K is acknowledged to be a sensitive, per-dataset tuned hyperparameter, a table of the chosen K, R, and other training details is required.","section":"Implementation Details; Methodology, 'Dual-Codebook for Discrete Representations'"},{"comment":"The paper does not explain how gradients are propagated through the hard argmin quantization in Eq. (1). If a straight-through estimator is assumed, it should be stated explicitly; if codebook vectors are updated only through the described 'interactive' update, that procedure should also be specified. Without this detail, the optimization of the encoder, decoder, and codebooks cannot be assessed.","section":"Eq. (1) and 'Dual-Codebook for Discrete Representations'"}],"minor_comments":[{"comment":"The text says AdaPoinTr and PoinTr achieved 6.10 and 6.26 respectively, but Table 2 lists their averages as 6.3 and 6.1; the order is reversed and should be corrected.","section":"Performance on ShapeNet_Part Dataset"},{"comment":"The two-paragraph description of the QIE ablation is duplicated almost verbatim; the duplicate should be removed.","section":"Ablation Study, 'Quantized Information Exchanging'"},{"comment":"The caption contains an extra 'and' in '(b) and the decoder-codebook'; it should be cleaned up.","section":"Figure 3 caption"},{"comment":"The reference for HyperCD (Zhu, Fan, and Weng 2024) appears to point to a survey on point cloud data augmentation, not to a point cloud completion method; the citation should be verified and corrected.","section":"References"},{"comment":"The summation in Eq. (4) uses indices starting at i=0, whereas the deduplicated sets were earlier defined with indices starting at t=1 and l=1; the notation should be made consistent.","section":"Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The central derivation and ablation logic are internally consistent, and the code release is a positive sign. The decisive issue is empirical support: the documented margins are small, single-run, and the codebook sizes are admitted to be tuned without a stated validation protocol. I would like to see variance estimates and a clean validation-based selection procedure before this can be accepted, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DC-PCN is a clean, incremental extension of VQ-based point cloud completion—separate encoder/decoder codebooks plus a quantized exchange mechanism—and the paper gives it a fair empirical shot. The main deliverables are the architecture, the ablations (including the shared-codebook failure case), and a public code link. The soft spot is exactly where the reader's report puts it: the SOTA claim rests on per-dataset codebook sizes chosen as 'optimal values' without a reported search protocol, and the reported margins are small enough that single-run numbers and selection effects could matter. That's a real concern, but it's a rigor problem, not a correctness problem.\n\nWhat's new: most VQ completion methods use a single codebook. Two codebooks at different depths, with a re-targeting/merging mechanism, is a plausible way to keep both structural and detail information. Group F in Table 5 is the most informative experiment—sharing one codebook across both levels hurts clearly (–0.15 CD-l1), so the dual design is doing something. The loss design (internal/external contrastive terms) is sensible and the paper describes it precisely enough to reimplement.\n\nSoft spots, in order of severity. First, the codebook-size search: the Limitations say the sizes are 'optimal values obtained by changing different parameters,' but the search space, criterion, and whether the test set was used are never stated. Since the PCN gain over AdaPoinTr is 0.07, the same size as the full method's gain over no-codebook, that margin is exactly the kind of number tuning can create. The authors need to report a validation split and a sweep table. Second, all results are single runs with no error bars or seed variance. Given the 0.01-level ablation contributions, run-to-run noise of that magnitude would erase the QIE and decoder-codebook ablations. Third, the paper cites VQ-DCTr but never compares against it—a missing baseline for a direct VQ-completion competitor. Minor editorial issue: the ablation section repeats a paragraph verbatim; that should be cleaned up.\n\nThe math and loss definitions look consistent; the evaluation uses standard external benchmarks and metrics, so no circularity. The central claim ('two codebooks help') holds up; the central SOTA claim is plausible but under-supported. A serious referee should ask for (1) a validation protocol for codebook sizes, (2) multiple seeds, and (3) the VQ-DCTr comparison. I'd send it to review—conditional accept territory, not desk reject. I'd put it on the reading-group list for anyone in 3D completion.","headline":"Solid incremental VQ-completion paper whose SOTA claim needs a validation protocol and variance estimates before the 0.07 margin over AdaPoinTr is believable.","tokens_in":13921,"tokens_out":2047,"would_cite":true,"duration_ms":21556,"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":"Two complementary codebooks give state-of-the-art point cloud completion accuracy.","keywords":["point cloud completion","dual-codebook quantization","vector quantization","transformer encoder-decoder","information exchange","Chamfer Distance","F-Score","3D shape reconstruction"],"falsifier":"A decisive experiment is to fix one codebook size for all datasets, select it only on a validation split, and retrain both DC-PCN and the closest baseline several times with different seeds. If the PCN CD-$\\ell_1$ gap of 0.07 shrinks or reverses, the state-of-the-art claim rests on per-dataset tuning rather than the method. A second check would compare the distances between quantized codes from two different samplings of the same surface against the distances between the unquantized features; if quantization does not pull same-surface codes closer, the invariance story is not supported.","tokens_in":12945,"feed_emoji":"🧊","tokens_out":10578,"duration_ms":99557,"temperature":0.7,"pith_summary":"Point cloud completion reconstructs a full 3D shape from a partial scan, and the paper argues that a core obstacle is ambiguity: the same object surface can be sampled into many different point clouds, so continuous features learned from one sampling can be inconsistent. DC-PCN tackles this by quantizing shallow and deep features against two codebooks, mapping similar surface features onto the same discrete code and thereby giving one surface one representation. The paper also contributes a quantized information exchange mechanism that deduplicates, re-targets, and merges codes between the two codebooks. On the PCN, ShapeNet_Part, and ShapeNet34 benchmarks, the authors report the best average Chamfer Distance and F-Score among all compared methods, including an average CD-$\\ell_1$ of 6.46 in the standard scaled units and an F-Score@1% of 0.850 on PCN. A sympathetic reader would care because a representation that is stable across samplings should produce cleaner, more faithful completions in downstream 3D tasks.","feed_headline":"Two codebooks beat previous bests in point cloud completion","feed_subtitle":"Quantizing shallow and deep features, DC-PCN tops Chamfer distance and F-Score on three benchmarks.","key_machinery":"The load-bearing object is the dual-codebook quantization scheme along with its quantized information exchange (QIE). An encoder-codebook $C_E$ and a decoder-codebook $C_D$, each a set of $K$ code vectors, replace each feature vector with its nearest code under Euclidean distance, following the VQ-VAE discretization idea. QIE has three parts: code deduplication, which removes repeated codes; code distribution re-targeting, which uses cascading MLPs to project codes from one codebook's distribution into the other's; and code merging, which blends matched codes by an adaptive weight equal to their cosine similarity. The loss combines Chamfer Distance on the coarse and complete clouds with an internal codebook loss that spreads codes apart and an external loss that pulls re-targeted codes toward their decoder-codebook matches.","core_discovery":"The central claim is that discretizing features with two complementary codebooks removes sampling-induced ambiguity and that this is what pushes completion quality past prior methods. The authors' argument is that point-based completion networks operate on continuous latent features, so different samplings of the same surface yield different features and the decoder must accommodate that variability. DC-PCN forces shallow features through an encoder-codebook and deep features through a decoder-codebook, replacing each feature vector with its nearest code vector. A quantized information exchange mechanism then deduplicates the two code sets, re-targets encoder codes into the decoder-codebook's distribution, and merges matched codes using a cosine-similarity weight. The paper reports top results on all three benchmarks, and on PCN the average CD-$\\ell_1$ of 6.46 and F-Score@1% of 0.850 are the best among the methods compared, with an MMD of 0.373 on KITTI.","pith_inferences":["Inference: the paper's own ablation does not separate the contribution of quantization from the contribution of the re-targeting MLP; isolating the two would test whether the key mechanism is discretization or simply extra learned parameters.","Inference: because the codebook sizes were tuned per dataset, an adaptive, data-driven codebook-size policy is the natural next step and would determine whether the method can be deployed on new scan distributions without per-dataset search.","Inference: the invariance rationale transfers beyond completion; any task in which multiple partial observations of one 3D surface must share a representation could be reframed as a dual-codebook quantization problem.","Inference: the reported margins over the closest baseline are small (0.07 in PCN CD-$\\ell_1$), so an independent multi-seed reproduction with rules fixed in advance would be the cleanest check that the advantage is systematic rather than tuning-dependent."],"forward_implications":["If the reported numbers reproduce, DC-PCN sets the best published results among the compared methods on the PCN benchmark, with CD-$\\ell_1$ 6.46 and F-Score@1% 0.850.","The ablations support the dual-codebook design: dropping the encoder-codebook costs 0.05 in CD-$\\ell_1$, dropping the decoder-codebook costs 0.01, and replacing both with a shared codebook costs 0.15.","On ShapeNet34's 21 unseen categories, DC-PCN still achieves the best CD-$\\ell_2$ among compared methods, which the paper reads as evidence of generalizability.","On real-world KITTI car scans, the method reports MMD 0.373, an improvement of 0.019 over the strongest compared baseline."],"supporting_citations":[{"why":"Supplies the VQ-VAE discrete representation learning mechanism that motivates codebook quantization.","marker":"Van Den Oord, Vinyals et al. 2017"},{"why":"ShapeFormer provides the codebook-based shape completion prior that the dual-codebook design is inspired by.","marker":"Yan et al. 2022"},{"why":"AutoSDF supplies an autoregressive codebook prior for 3D completion, another source of the codebook idea.","marker":"Mittal et al. 2022"},{"why":"Introduces the PCN dataset and point-based completion network that serve as the main benchmark and early baseline.","marker":"Yuan et al. 2018"},{"why":"AdaPoinTr is the closest competing baseline whose numbers DC-PCN aims to beat on PCN, ShapeNet34, and KITTI.","marker":"Yu et al. 2023"},{"why":"Defines the F-Score@1% metric used alongside Chamfer Distance in the evaluations.","marker":"Tatarchenko et al. 2019"},{"why":"SeedFormer supplies a patch-seed transformer baseline and comparison method for the benchmark tables.","marker":"Zhou et al. 2022"}],"fun_headline_variants":["Dual-codebook quantization lifts point cloud completion to SOTA","Quantized features fix sampling ambiguity in 3D completion","Two codebooks deduplicate features for better point cloud completion","DC-PCN: Codebook quantization reduces sampling variability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-dataset codebook sizes, which the authors tuned and report as optimal, were selected without using test-set labels; if they were tuned to minimize test error, the small reported margins over the closest baseline could be a selection artifact rather than a property of the method.","fun_headline_variants_meta":{"raw":{"variants":["Dual-codebook quantization lifts point cloud completion to SOTA","Quantized features fix sampling ambiguity in 3D completion","Two codebooks deduplicate features for better point cloud completion","DC-PCN: Codebook quantization reduces sampling variability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1474,"prompt_tokens":968,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":438}},"tokens_in":584,"tokens_out":506,"duration_ms":5784,"temperature":1.0,"reasoning_tokens":438,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:45:57.727539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment is to fix one codebook size for all datasets, select it only on a validation split, and retrain both DC-PCN and the closest baseline several times with different seeds. If the PCN CD-$\\ell_1$ gap of 0.07 shrinks or reverses, the state-of-the-art claim rests on per-dataset tuning rather than the method. A second check would compare the distances between quantized codes from two different samplings of the same surface against the distances between the unquantized features; if quantization does not pull same-surface codes closer, the invariance story is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the VQ-VAE discrete representation learning mechanism that motivates codebook quantization."},{"cited_title":"J.; Lischinski, D.; Cohen-Or, D.; and Huang, H","cited_arxiv_id":null,"evidence_quote":"ShapeFormer provides the codebook-based shape completion prior that the dual-codebook design is inspired by."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AutoSDF supplies an autoregressive codebook prior for 3D completion, another source of the codebook idea."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the PCN dataset and point-based completion network that serve as the main benchmark and early baseline."},{"cited_title":"R.; Ranftl, R.; Li, Z.; Koltun, V.; and Brox, T","cited_arxiv_id":null,"evidence_quote":"Defines the F-Score@1% metric used alongside Chamfer Distance in the evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SeedFormer supplies a patch-seed transformer baseline and comparison method for the benchmark tables."}],"review_version":1}