{"id":"370fec21-0e46-425d-b247-99306167ac55","arxiv_id":"2502.01855","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A fine-to-coarse training scheme for cuboid shape abstraction achieves more accurate and more compact 3D shape representations than prior cuboid-based methods.","lead":"This paper presents an unsupervised method that first reconstructs a 3D shape with many cuboids, then gradually reduces them to a compact set of a few cuboids that still capture the shape. The authors report improved reconstruction accuracy and volume preservation over previous cuboid-based approaches, with potential uses in shape segmentation, retrieval, and symmetry detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Volume loss depends on an undocumented volumetric ground-truth point cloud X_V; without it the method's IoU advantage collapses (Table 3), so the central claim is not independently reproducible.","rationale":"The reader's weakest-assumption identification is the same one I regard as most load-bearing. The ablation in Table 3 isolates L_vol as the component that prevents degenerate, surface-thin cuboids and produces the reported IoU; L_vol requires X_V, a volumetric point cloud of the target shape. The paper never specifies how X_V is constructed for the ShapeNet and DFAUST benchmarks, so the primary contribution cannot be reproduced or even fully assessed from the text. This is not an internal inconsistency, but a missing support step in the central argument: the claimed improvement over prior work rests on a data-preprocessing detail that is absent.\n\nOther concerns in the reader's verdict are legitimate but secondary. The best-of-five baseline selection, per-class merge-threshold tuning, and absence of error bars affect the confidence intervals around the reported margins; the abstract's 'fewer cuboid primitives' phrasing is contradicted by Table 1 for chair and table, but the paper's Section 4.1 already qualifies this as 'least or second least'. None of these threaten the core mechanism in the same way as the undocumented X_V dependency: if X_V is generated differently, the volume loss changes, and with it the central numbers.\n\nBecause the method itself is plausible and the concern is about missing reproducibility rather than a demonstrated flaw, I keep the reader's CONDITIONAL verdict. The paper should be accepted only after the authors release the exact volumetric sampling procedure and demonstrate that the reported CD/IoU values are reproducible under that procedure.","tokens_in":13154,"tokens_out":5838,"duration_ms":59481,"concrete_test":"Request the X_V generation code and retrain the plane-class baseline (L_rec = L_surf + L_vol, lambda_abs = 1e-3, half-cos) with the exact procedure; if code is unavailable, implement uniform interior sampling from watertight meshes (e.g., 50k points per shape via signed-distance rejection sampling) and rerun the same configuration. The concern lands if Table 3 baseline CD/IoU (0.024/61.1) and Table 1 plane CD/IoU (0.026/56.0) cannot be reproduced within a small tolerance, or if the L_rec = L_surf ablation does not yield IoU approximately 0.0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim ('represent 3D shapes more precisely with fewer cuboid primitives than previous work', abstract) is carried by the volume-preservation loss L_vol (Eq. 6). Section 3 states the network is trained with 'a surface point cloud X_S and a volume point cloud X_V', but Section 4 ('Datasets') never specifies how X_V is generated for ShapeNet or DFAUST, nor whether meshes are watertight before sampling. This is load-bearing because Table 3's ablation shows that removing L_vol (L_rec = L_surf) drops IoU to 0.0 and degrades CD from 0.024 to 0.058 on the plane class; the same Table attributes the avoidance of degenerate thin cuboids to L_vol. If X_V is unavailable (e.g., surface-only scans), the method cannot be used as described, and the paper's Limitations section (4.6) does not disclose this dependency. The reported IoU gains are therefore not reproducible or assessable without the missing X_V generation procedure, and the 'unsupervised' label is misleading in that training requires volumetric occupancy ground truth, even though part labels are not used.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fine-to-coarse unsupervised learning method for cuboid shape abstraction. The network is trained to predict a large set of cuboids and then gradually reduces their number through a loss that penalizes redundant primitives, subject to a user-specified target count. A second contribution is a volume-preserving reconstruction loss that complements the surface Chamfer loss. The method is evaluated on ShapeNet classes (plane, chair, table) and a DFAUST subset, where it reports improved Chamfer distance and IoU relative to HCA, CAS, and DPF-PPM, often with fewer cuboids. The abstraction is also applied to co-segmentation, clustering, retrieval, and partial symmetry detection.","tokens_in":13443,"tokens_out":4137,"duration_ms":40679,"significance":"If the results hold, the fine-to-coarse training schedule is a worthwhile alternative to fixed-primitive-count optimization, and the explicit volume-preservation loss directly addresses a known degeneration of cuboid abstractions (thin surface shells). The paper provides user-controlled compactness and demonstrates downstream utility. It is, however, an empirical paper without code or released models, and its central claims rest on an undocumented volumetric supervision signal and on a benchmark protocol that does not fully report variance. These issues are fixable and do not invalidate the approach, but they must be addressed before the state-of-the-art claims can be assessed.","major_comments":[{"comment":"The volume point cloud X_V is never defined. The paper states in Section 3 that the network is trained with 'a surface point cloud X_S and a volume point cloud X_V', but Section 4 does not specify how X_V is generated for ShapeNet or DFAUST, how many interior points are sampled, whether meshes are made watertight beforehand, or whether this information is available at inference time. This is load-bearing: Table 3 shows that removing L_vol (L_rec = L_surf) collapses IoU to 0.0 and increases CD from 0.024 to 0.058 on the plane class, and the paper attributes the avoidance of degenerate thin cuboids to L_vol. The claim that the method is 'unsupervised' is also overstated, because training requires volumetric occupancy ground truth (interior samples) even though part labels are not used. Section 4.6 does not disclose this dependency. Please provide the exact X_V generation procedure, state its availability requirements, and either re-word 'unsupervised' or justify it as unsupervised with respect to semantic labels.","section":"Section 3.2.1 (Eq. 6) and Section 4 (Datasets)"},{"comment":"The comparison protocol is asymmetric. For the baselines, the paper states (Section 4, Reference methods): 'we train the method five times for every category using the standard hyper-parameters and report the best performing run in terms of Chamfer distance.' For the proposed method, no such multi-seed protocol is reported for the main results in Table 1, and no variance or significance tests are given. On several key entries the reported advantage is small or zero (e.g., chair CD: Ours 0.036 vs CAS 0.036; chair Num: Ours 8.37 vs CAS 9.77, but the CD tie means the 'more precise' claim is not supported there). The paper should report mean and standard deviation over at least five runs for all methods, or clearly state that the same best-of-five selection was applied to both the baselines and the proposed method.","section":"Section 4 (Reference methods) and Section 4.1 (Table 1)"},{"comment":"The merge threshold theta_merge is a per-class tuned post-processing parameter (plane 1.2, chair 1.4, table 1.0, human 1.0). Table 4 shows that on the plane class, increasing theta_merge monotonically improves both CD and IoU, and that the 'w/o' row (no merging) gives the best CD (0.024) and IoU (61.1) of all configurations. Thus the reported 'best scores ... while using least cuboids' configuration is obtained by deliberately sacrificing reconstruction quality to reduce primitive count, and the choice of theta_merge is not derived from a principled criterion. The paper should either present the full trade-off curve and recommend a criterion, or show that the chosen theta_merge is stable across classes and random seeds rather than selected per benchmark.","section":"Section 3.3 and Table 4"}],"minor_comments":[{"comment":"The caption says 'Intersection of Union' but should be 'Intersection over Union'.","section":"Section 4.1, Table 1 caption"},{"comment":"There is a typo: 'learnarble cuboid latents' should be 'learnable cuboid latents'.","section":"Section 4.5"},{"comment":"The captions use 'distinc parts' instead of 'distinct parts'.","section":"Figures 5 and 6 captions"},{"comment":"The notation gamma_bar_m is confusing: the sentence 'gamma_bar_m is a shorthand notation to denote that there exists a primitive closer than primitive m' does not align with the product over m-bar < m in Eq. (3). Please clarify the ordering and the exact meaning of gamma_bar_m.","section":"Section 3.2.1, Eq. (3)"},{"comment":"The row 'w/o l_c' reports Num=37.32 and moderate CD/IoU, but the text says 'training without using l_c explicitly diverges.' If the configuration diverges, it would be clearer to state that the reported metrics correspond to the best of five runs that did not diverge, or to omit the row and simply state the failure.","section":"Section 4.5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising and the ablation study is informative, but the undocumented volumetric supervision signal (X_V) is a serious reproducibility gap that must be closed before the paper can be accepted. I also recommend the editor ask the authors to release code and data-generation scripts, since without them the exact X_V sampling procedure is unlikely to be reproduced reliably. The best-of-five baseline selection without equivalent reporting for the proposed method is a fairness issue that should be fixed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful paper that deserves refereeing, and the main caveat is exactly where the stress-test points — the volume point cloud X_V is never specified.\n\nWhat's new and good: unlike HCA (coarse-to-fine), CAS, and Tulsiani (fixed counts), the network starts with 128 cuboids and smoothly prunes to ~7–10 during training. The abstraction loss with binary cross-entropy on existence probabilities is a clean way to steer the count. The volume-preservation loss L_vol is new in this cuboid setting, and the ablation (Table 3) makes a convincing case that it is the difference between IoU 61 and IoU 0, and between sensible cuboids and degenerate slivers. The merging post-process is straightforward, and the threshold analysis (Table 4) honestly shows the trade-off. Reported CD/IoU numbers are directionally better than CAS/HCA at comparable or lower primitive counts on most classes. This is a genuine step for compact structural abstraction.\n\nSoft spots, in proportion. The missing X_V generation is the biggest. The method needs volumetric point clouds for training, but Section 4 never says how they are obtained — mesh voxelization, interior sampling, watertightness? Table 3 shows removing L_vol collapses IoU to 0.0, so the headline advantage rides on it. The Limitations section does not mention this dependency. That is a reproducibility hole, not a fatal flaw, but it needs to be fixed by specifying the procedure or releasing code and data.\n\nStatistics are thin. Baselines are selected as best-of-five runs; the authors do not report variance for their own runs, and the merge threshold is tuned per class. The abstract says “fewer cuboids than previous work,” but Table 1 shows they use fewer than CAS and DPF while using more than HCA on plane and chair. That overstatement should be toned down. Also, the “unsupervised” label is fair only in the sense that no part labels are used; training requires volumetric occupancy ground truth, which is a form of dense annotation. A sentence of clarification would help.\n\nMinor: the downstream sections (clustering, retrieval, symmetry) are qualitative demonstrations — fine for a graphics paper but not evidence.\n\nWho this is for: anyone working in shape abstraction, structural co-segmentation, or primitive-based reconstruction. The method is practical, the writing is clear, and the ablations are informative. It deserves peer review; a careful referee should ask for the X_V procedure, error bars, and code. I would bring it to reading group in current form, mostly to discuss what “unsupervised” really means here.\n\nRecommendation: engage, with revision expectations.","headline":"Fine-to-coarse cuboid abstraction with a volume-preservation loss is a genuinely new and effective recipe, but the reported gains lean on an undocumented volumetric ground truth and on thin statistics.","tokens_in":13951,"tokens_out":1927,"would_cite":true,"duration_ms":19586,"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":"A fine-to-coarse training schedule with volume preservation yields 3D cuboid abstractions that are both more compact and more accurate than previous methods.","keywords":["shape abstraction","3D structural representation","shape co-segmentation","cuboid primitives","fine-to-coarse training","volume preservation","point cloud","unsupervised learning"],"falsifier":"Train the same architecture and schedule on a dataset that offers only surface point clouds (no interior samples) and evaluate IoU on the test set; the paper's own ablation predicts that the cuboids degenerate to thin plates and IoU collapses toward zero. If the method still maintains high IoU under purely surface supervision, the central claim about volume preservation would be overturned.","tokens_in":12949,"feed_emoji":"📦","tokens_out":7909,"duration_ms":61814,"temperature":0.7,"pith_summary":"This paper claims that cuboid abstraction of 3D shapes improves when the network starts with a deliberately overcomplete set of primitives and gradually prunes them during training, rather than fitting a fixed small number from the outset. The authors add a volume-preservation loss on top of the usual surface Chamfer loss, which prevents thin degenerate cuboids and improves intersection-over-union with the original volume. On ShapeNet and DFAUST categories, they report the best or second-best Chamfer distance and IoU while using the fewest cuboids among cuboid-based baselines. If correct, unsupervised structural abstraction becomes accurate enough to support co-segmentation, retrieval, and symmetry detection without manual annotations.","feed_headline":"Fine-to-coarse cuboid pruning beats prior 3D shape abstraction","feed_subtitle":"Prune from 128 cuboids to a few for better fidelity and fewer primitives in 3D abstraction.","key_machinery":"The load-bearing mechanism is the fine-to-coarse schedule, implemented by an abstraction loss that forces the number of active cuboids toward a prescribed target via binary cross-entropy on existence probabilities, plus a reconstruction loss that combines surface Chamfer distance with a volume-preservation Chamfer term sampled from cuboid interiors. The architecture processes the input point cloud through two Vision Transformers: the first builds global shape features from local point patches, and the second lets learnable cuboid latents attend to those features to predict each cuboid's rotation, translation, scale, and existence probability. During inference, a merging post-process replaces highly overlapping cuboids with the oriented bounding box of their union, further reducing the primitive count without substantial loss of accuracy.","core_discovery":"The central claim is that a fine-to-coarse training scheme, combined with a reconstruction loss that enforces volume preservation as well as surface approximation, produces cuboid shape abstractions that are simultaneously more compact and more faithful than those of prior work. The network begins with 128 cuboids and is driven to a user-specified minimum count through a binary cross-entropy abstraction loss applied to each cuboid's existence probability; this lets it lock in fine details early and then discard or merge redundant primitives. A volume loss, computed against interior point samples, prevents the degenerate thin cuboids that surface-only losses allow. In the plane, chair, table, and human categories, the method achieves the best or near-best Chamfer distance and IoU while using the least or second-least number of cuboids, and the resulting cuboids support co-segmentation, clustering, retrieval, and partial symmetry detection.","pith_inferences":["The method's reported gains depend on volumetric point clouds for the volume loss, which are not available from surface-only scans; a practical deployment would need a way to estimate or predict interior points.","The merge post-processing could be reformulated as a differentiable loss or a learned selection module, allowing the network to predict the final compact abstraction in a single forward pass.","The fine-to-coarse schedule resembles a curriculum on primitive count; the same principle might improve other structured prediction tasks such as skeleton extraction or part-graph discovery.","Since the paper notes that re-indexing merged cuboids could improve co-segmentation, the current indexing scheme likely underestimates the method's semantic consistency."],"forward_implications":["Unsupervised cuboid abstraction can reach state-of-the-art surface and volume fidelity while using fewer primitives, making the representation more useful for structure-aware modeling and editing.","The fine-to-coarse schedule is a transferable strategy for primitive-fitting networks and could extend to superquadrics or deformable templates by changing the primitive parameterization.","Because inference needs only a surface point cloud, the trained model can abstract new shapes on demand without volumetric supervision.","The learned cuboids carry enough structural signal to support co-segmentation, clustering, retrieval, and partial symmetry detection without annotation."],"supporting_citations":[{"why":"Supplies the Chamfer-distance reconstruction loss formulation (Equations 1–4) that the paper extends with a volume term.","marker":"[Paschalidou et al. 2019]"},{"why":"Provides the cuboid decoder design and the main cuboid-based baseline (CAS) that the method builds on and compares against.","marker":"[Yang and Chen 2021]"},{"why":"Defines the hierarchical cuboid abstraction baseline (HCA) that the fine-to-coarse approach is contrasted with.","marker":"[Sun et al. 2019]"},{"why":"Supplies the deformable primitive field baseline (DPF) whose primitive prediction module is adapted for cuboid comparison.","marker":"[Shuai et al. 2023]"},{"why":"Provides the point cloud encoder architecture that produces local shape features for the network.","marker":"[Zhang et al. 2022]"},{"why":"Contributes the ShapeNet categories (plane, chair, table) used for training and evaluation.","marker":"[Chang et al. 2015]"},{"why":"Contributes the DFAUST human shapes dataset used for the human category benchmark.","marker":"[Bogo et al. 2017]"}],"fun_headline_variants":["Fine-to-coarse cuboids: fewer primitives, better 3D fit","Cuboid abstraction improved with volume-aware pruning","From 128 to a few cuboids in unsupervised 3D abstraction","Fine-to-coarse training yields compact, accurate cuboids","Pruning cuboids fine-to-coarse improves 3D shape abstraction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that volumetric point clouds of the training shapes are available to compute the volume-preservation loss; the paper never specifies how these interior samples are generated, and removing the volume loss drops IoU to zero in the ablation.","fun_headline_variants_meta":{"raw":{"variants":["Fine-to-coarse cuboids: fewer primitives, better 3D fit","Cuboid abstraction improved with volume-aware pruning","From 128 to a few cuboids in unsupervised 3D abstraction","Fine-to-coarse training yields compact, accurate cuboids","Pruning cuboids fine-to-coarse improves 3D shape abstraction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000919,"raw_usage":{"total_tokens":3940,"prompt_tokens":939,"completion_tokens":3001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2911}},"tokens_in":555,"tokens_out":3001,"duration_ms":20897,"temperature":1.0,"reasoning_tokens":2911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:13:04.253491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture and schedule on a dataset that offers only surface point clouds (no interior samples) and evaluate IoU on the test set; the paper's own ablation predicts that the cuboids degenerate to thin plates and IoU collapses toward zero. If the method still maintains high IoU under purely surface supervision, the central claim about volume preservation would be overturned.","supporting_citations":[],"review_version":1}