{"id":"34a46a7d-92ea-47ce-97c2-76faa34de9eb","arxiv_id":"2504.13788","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"RefComp completes partial 3D point clouds by fusing features of retrieved reference shapes into the input's latent representation, and it can be trained jointly across object classes.","lead":"This paper introduces RefComp, a method that completes partial 3D point clouds without paired ground-truth data by retrieving similar complete shapes as references. It claims the first class-agnostic version that works across object categories and reports state-of-the-art results on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Class-agnostic claim rests on test-time same-class reference retrieval, which the stated protocol requires but does not justify.","rationale":"The reader identified exactly the load-bearing weakness: the class-agnostic protocol still relies on same-class reference retrieval at test time. I find no additional concern that changes the picture. The method itself is coherent and the class-aware results appear competitive, but the strongest claim about class-agnostic generalization is not established by the current experimental setup. Since the reader already assigned a conditional verdict based on this issue, my stress-test confirms that verdict without moving it. The appropriate fix is either a label-free retrieval evaluation or a clear disclosure that class labels are required for reference selection.","tokens_in":20979,"tokens_out":3929,"duration_ms":38876,"concrete_test":"Implement a label-free retrieval variant for the RefComp Unified evaluation: for each test partial px, choose the reference pair minimizing Chamfer Distance over the entire ShapeNet database with no class filter, then rerun the 3D-EPN/CRN and real-world tables. Compare CD/F1/UCD/MMD against the published same-class-retrieval numbers and report the cross-class selection rate. If performance degrades substantially or cross-class references dominate, the claim should be revised to 'single model with class-conditional retrieval.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RefComp is the first unified framework for class-agnostic unpaired point cloud completion. The evidence does not support that claim as stated. In Section III-A, reference pairs are selected by the Chamfer Distance between px and all py 'of the same class.' In Section IV.C, testing 'only use[s] the reference pair that is the closest to px of the same class.' No mechanism is described for obtaining class labels at inference or for filtering the reference database without labels. Thus the RefComp Unified results in Tables I, II, IV, and V are produced with a class-conditional retrieval protocol, so the 'class-agnostic' claim is not supported as written. If labels are unavailable, or if the database lacks a same-class shape similar to px, the reported CD/F1/UCD/MMD numbers cannot be reproduced. The ablation study and parameter-sharing design are plausible, but the headline generalization claim depends on an unstated assumption about test-time label availability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RefComp, a reference-guided framework for unpaired point cloud completion. The method builds partial-complete reference pairs by degrading complete shapes from a database using the target partial cloud as a corruption template, then selects the top-N pairs by Chamfer distance. A reference branch and a target branch share parameters; a Latent Shape Fusion Module (LSFM) fuses missing-region mask features into the partial latent code, and the decoder produces the completed cloud. Training losses include Chamfer reconstruction on the reference branch, degradation-consistency on the target branch, and a Wasserstein loss that aligns the completed latent features with the complete latent space. The paper reports class-aware and class-agnostic training variants ('RefComp' and 'RefComp Unified') and evaluates them on CRN, 3D-EPN, PartNet, KITTI, ScanNet, and MatterPort3D, with ablations on the reference-selection strategy, the LSFM module, and parameter sharing.","tokens_in":21155,"tokens_out":6806,"duration_ms":68886,"significance":"If the class-agnostic claim were fully supported, RefComp would be a practically valuable single-model solution for unpaired completion across many object classes, and the reference-guided latent translation formulation is a sensible way to inject shape priors without paired ground truth. The paper gives credit for a clear formulation, extensive comparisons, and ablations that show the parameter-sharing mechanism and the LSFM are important for the reported performance. However, the evaluation protocol does not currently support the headline 'class-agnostic' claim, because reference retrieval at test time uses the same-class oracle. The central technical machinery is plausible, but the gap between the claimed contribution and the experimental protocol needs to be closed before the paper can be accepted as stated.","major_comments":[{"comment":"The central 'class-agnostic' claim is not supported by the evaluation protocol. Section III-A states that reference pairs are selected by Chamfer distance between px and all py 'of the same class,' and Section IV.C states that at test time 'we only use the reference pair that is the closest to px of the same class.' Consequently, every 'RefComp Unified' entry in Tables I, II, IV, and V is obtained with a same-class oracle for retrieval; the model is not evaluated under the label-free conditions that the introduction uses to motivate class-agnostic completion. To make the claim stand, the authors should either (a) evaluate with a class-label-free retrieval protocol over a pooled multi-class reference database, (b) report both oracle and label-free numbers, and (c) discuss the case where no same-class reference similar to px exists in the database. As written, the method is best described as a single model trained across classes with class-conditional retrieval, not a class-agnostic completion pipeline.","section":"Section III-A and Section IV.C"},{"comment":"The description of the 'class-agnostic training setting' (RefComp Unified) is incomplete. The paper does not state whether the reference pairs used during Unified training are also selected by same-class Chamfer distance (as in Section III-A) or by a label-free rule. If same-class selection is used during training, then class labels are required at train time for every reference pair, which further weakens the 'unified/class-agnostic' claim; if label-free selection is used, the retrieval rule and any changes to the Section III-A pipeline should be specified. This distinction is necessary because the loss and the pseudo-label role of the reference data depend on which pairs are admitted.","section":"Section III-A and Section IV.C"}],"minor_comments":[{"comment":"In Eq. (3), the notation defines '+' as element-wise addition, but the prose reads 'denoted element-wise addition'; the grammar should be corrected, and the precedence of concatenation versus addition in the fused expression should be clarified.","section":"Equation (3)"},{"comment":"The text states 'we use a minimum CD = 1.0' when selecting reference pairs, but the CD values in the tables are scaled by 1e4; please specify whether the threshold is also scaled and give the threshold in the same units as the reported metrics.","section":"Section IV.A"},{"comment":"The entries 'Generalized + Random', 'Generalized + Matching', 'Fine-tuned + Random', and 'Fine-tuned + Matching' are not explicitly mapped to RefComp variants in the text; please add a sentence explaining what each configuration is and how it relates to RefComp Unified.","section":"Table VI"},{"comment":"The complete-shape database used for reference retrieval is described as ShapeNet with 55 classes, but it is not stated whether the same database is used for the KITTI, ScanNet, and MatterPort3D experiments; please specify the reference database used for each real-world benchmark.","section":"Section III-A and Section IV.A"},{"comment":"Several references are duplicated under different numbers (e.g., [16]/[39], [20]/[40], and [22]/[63]); please consolidate the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid internal architecture and the ablations are informative, but the same-class retrieval at test time is a significant gap between the stated contribution and the experiments. Adding a label-free retrieval evaluation, or substantially revising the claims to reflect class-conditional reference selection, is essential before this can be considered for acceptance. The novelty claim of being the 'first unified framework for class-agnostic unpaired point cloud completion' should also be qualified in light of existing reference-guided and self-supervised methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RefComp is a genuine incremental advance: it casts unpaired point cloud completion as reference-guided shape translation in latent space, and the Latent Shape Fusion Module with parameter sharing is a reasonable, well-ablated design. The class-aware numbers are competitive, the ablation study supports the key modules, and the real-world results are interesting if you trust the metrics. The reference-branch supervision through self-generated partial-complete pairs is legitimate self-supervision, not circular in the way the stress-test note worried about.\n\nThe soft spot is exactly where the skeptics pointed. Section III-A selects reference pairs by Chamfer distance 'of the same class,' and Section IV.C says test-time retrieval uses the reference closest to px 'of the same class.' That means the class-agnostic evaluation leaks class labels into the retrieval stage. The model parameters might be shared across classes, but the pipeline as evaluated is class-conditional. So the 'first unified class-agnostic framework' claim is overstated as written. The gap is addressable: either run retrieval without labels (nearest neighbor in shape space, no class filter) or explicitly disclose that class labels are needed at inference. If the latter, the 'class-agnostic' label should be softened to 'class-agnostic training with class-conditional retrieval.'\n\nThe lack of error bars and released code makes the results harder to verify, but that's a common weakness and not a fatal one. The citation pattern looks fine, and the related work is covered adequately.\n\nThis paper deserves a serious referee. The central idea is plausible, the experiments are extensive, and the flaw is specific and fixable. I would reject the current version only because the headline claim is not supported; a revision with label-free retrieval results or a more honest claim could change my mind. I'd cite it if the code comes out, and I'd bring it to a reading group focused on 3D vision or self-supervised learning, with the caveat that the class-agnostic headline needs scrutiny.","headline":"A solid reference-guided completion framework whose headline class-agnostic claim is undercut by test-time same-class retrieval; worth refereeing but needs a label-free evaluation or a revised claim.","tokens_in":21700,"tokens_out":1406,"would_cite":false,"duration_ms":15292,"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":"RefComp claims to be the first unified framework for class-agnostic unpaired point cloud completion, reformulating completion as reference-guided shape translation in latent space.","keywords":["unpaired point cloud completion","class-agnostic completion","reference-guided shape translation","latent shape fusion","retrieval-based completion","Chamfer distance","3D shape completion"],"falsifier":"Hold out all complete shapes of one class, say cabinets, from the reference database while training the class-agnostic model, then test on partial cabinets from that class. If completion quality on the held-out class falls to the level of a random-reference baseline, the class-agnostic claim is falsified; if it stays high, the model is learning a generic shape prior rather than depending on same-class retrieval.","tokens_in":20774,"feed_emoji":"🧩","tokens_out":9708,"duration_ms":81632,"temperature":0.7,"pith_summary":"RefComp tries to establish that unpaired point cloud completion—filling in partial 3D scans without paired ground-truth data—does not need one model per object class. For each partial input, the framework retrieves a complete shape of the same class from a multi-class reference database, degrades that complete shape into a partial-complete pair, and uses the pair to guide completion in a shared latent space. The paper reports that this design is the first unified class-agnostic unpaired completion method, achieves state-of-the-art results in the class-aware training setting, and stays competitive in the class-agnostic setting on virtual-scan and real-world benchmarks. If this holds, a single trained model could replace per-class completion systems and remove the need to identify an object's class before completing it.","feed_headline":"One model completes any object class with no paired ground truth","feed_subtitle":"Unpaired reference-guided framework matches or beats per-class completion models on virtual and real scans.","key_machinery":"The load-bearing component is the Latent Shape Fusion Module (LSFM), a feature-fusion network with a reference branch and a target branch that share every parameter. It takes the encoded partial feature $z_{p_y}$ (or $z_{p_x}$) and a mask feature $z_{m_y}$, which is the embedding of the region missing from the reference partial (complete $c_y$ minus partial $p_y$), and produces completed features through residual blocks, concatenation, and element-wise additions. The mask is the carrier of the structural information: because the reference partial was corrupted using the target as a template, the mask tells the network where the target is likely missing points. A Wasserstein-distance loss on the reference branch pulls both completed feature codes into the complete-latent space, while Chamfer-Distance losses on the degraded reconstruction of the target and on the reconstruction of the reference supervise the decoder.","core_discovery":"The central claim is that unpaired completion can be recast as a shape translation problem: move the latent code of a partial point cloud into the latent space of complete shapes, using retrieved reference pairs as the only supervision. A reference pair is built by taking a complete cloud $c_y$ from a multi-class database, corrupting it using the target partial $p_x$ as a template to produce $p_y$, and selecting the top-N pairs by Chamfer distance. The Latent Shape Fusion Module then fuses the partial features with the missing-region mask embedding and, because the reference and target branches share parameters, the target branch inherits the reference branch's completion ability; a Wasserstein-distance loss pulls the completed features into the complete-latent space. The authors report state-of-the-art average Chamfer Distance and F1 in class-aware training over eight classes, and a class-agnostic version ('RefComp Unified') that outperforms several class-aware baselines, including on real-world KITTI, ScanNet, and MatterPort3D data.","pith_inferences":["The retrieval step selects same-class references by Chamfer distance, but the architecture never consumes class labels; a direct extension would drop the class filter and retrieve purely by geometry, testing whether similarity alone can carry class-agnostic completion on unlabeled scans.","Because the mask feature is computed as complete-minus-partial for the reference pair, the same fusion machinery could be driven at inference time by a learned occlusion estimator, making the framework applicable to arbitrary partiality patterns instead of KNN-corrupted ones.","The paper's ablation shows that randomly selecting among the top-3 retrieved references improves CD by about 2.0 over using one fixed reference, so retrieval diversity acts as a regularizer; a harder-negative or adversarial retrieval scheme is a natural next step.","A stress test implicit in the design is to grow the reference database over time, e.g., from a scanner's accumulated observations; if completion quality scales with database size, the framework would double as a lifelong completion system."],"forward_implications":["If the class-agnostic claim holds, one model trained on a pooled multi-class database can be deployed on arbitrary partial scans, removing per-class training, per-class storage, and inference-time class identification.","The reported KITTI inference times of 0.02 to 0.04 seconds per object with a 40.5-million-parameter model put the pipeline in a range usable for autonomous-driving perception.","On real-world scans from ScanNet and MatterPort3D, the class-aware versions report lower MMD without fine-tuning than supervised baselines that were fine-tuned to those datasets, indicating that the reference data transfers to depth-sensor noise.","Ablation shows that removing parameter sharing between the reference and target branches raises average Chamfer Distance on the CRN subset from 14.4 to 31.1, identifying the shared branches as the main carrier of the completion signal."],"supporting_citations":[{"why":"Supplies the non-parametric degradation module that corrupts a complete cloud into a partial one using the target partial as template, and the KITTI evaluation protocol.","marker":"[35]"},{"why":"Primary unpaired baseline to beat; also provides the discriminator design used in RefComp w/dis and the PartNet MMD protocol.","marker":"[32]"},{"why":"Unpaired cycle-consistency baseline for CD/F1 comparisons; its GAN architecture is used as the 'Only GAN' ablation.","marker":"[31]"},{"why":"First unpaired point cloud completion method; establishes the adversarial training setting and baseline numbers on real scans.","marker":"[30]"},{"why":"Self-supervised baseline P2C whose 3D-EPN sofa-replacement protocol is followed in the asterisk-marked experiments.","marker":"[38]"},{"why":"ShapeNet is the source of the complete point clouds used to create reference pairs and of the virtual-scan datasets.","marker":"[62]"},{"why":"PointNet is the two-layer encoder backbone used for the partial and complete encoders.","marker":"[14]"},{"why":"Supplies the Wasserstein-distance loss that aligns fused features with the complete latent space.","marker":"[61]"},{"why":"Diffusion-based unpaired baseline R2C compared on plane, chair, and table classes.","marker":"[34]"}],"fun_headline_variants":["Shape translation recasts unpaired completion, no ground truth","Class-agnostic point cloud completion with reference pairs","Latent space fusion enables unpaired completion for any class","One model, any object: reference-guided unpaired completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework works only if, for every partial scan to be completed, the reference database contains a complete shape of the same object class that is geometrically similar enough to be retrieved by Chamfer distance; if no such reference exists, or a wrong class is retrieved, the mask encodes the wrong missing region and the completion guidance fails.","fun_headline_variants_meta":{"raw":{"variants":["Shape translation recasts unpaired completion, no ground truth","Class-agnostic point cloud completion with reference pairs","Latent space fusion enables unpaired completion for any class","One model, any object: reference-guided unpaired completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000881,"raw_usage":{"total_tokens":3836,"prompt_tokens":1006,"completion_tokens":2830,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2764}},"tokens_in":622,"tokens_out":2830,"duration_ms":18428,"temperature":1.0,"reasoning_tokens":2764,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:00:04.899348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out all complete shapes of one class, say cabinets, from the reference database while training the class-agnostic model, then test on partial cabinets from that class. If completion quality on the held-out class falls to the level of a random-reference baseline, the class-agnostic claim is falsified; if it stays high, the model is learning a generic shape prior rather than depending on same-class retrieval.","supporting_citations":[{"cited_title":"Unsupervised 3d shape completion through gan inversion","cited_arxiv_id":null,"evidence_quote":"Supplies the non-parametric degradation module that corrupts a complete cloud into a partial one using the target partial as template, and the KITTI evaluation protocol."},{"cited_title":"Learning a structured latent space for unsupervised point cloud completion,","cited_arxiv_id":null,"evidence_quote":"Primary unpaired baseline to beat; also provides the discriminator design used in RefComp w/dis and the PartNet MMD protocol."},{"cited_title":"Cy- cle4completion: Unpaired point cloud completion using cycle transfor- mation with missing region coding","cited_arxiv_id":null,"evidence_quote":"Unpaired cycle-consistency baseline for CD/F1 comparisons; its GAN architecture is used as the 'Only GAN' ablation."},{"cited_title":"Unpaired point cloud completion on real scans using adversarial training,","cited_arxiv_id":null,"evidence_quote":"First unpaired point cloud completion method; establishes the adversarial training setting and baseline numbers on real scans."},{"cited_title":"P2c: Self-supervised point cloud completion from single partial clouds,","cited_arxiv_id":null,"evidence_quote":"Self-supervised baseline P2C whose 3D-EPN sofa-replacement protocol is followed in the asterisk-marked experiments."},{"cited_title":"Reverse2complete: Unpaired multimodal point cloud completion via guided diffusion,","cited_arxiv_id":null,"evidence_quote":"Diffusion-based unpaired baseline R2C compared on plane, chair, and table classes."}],"review_version":1}