{"id":"9e862ed6-4094-45e4-86be-7c9e46e53f1b","arxiv_id":"2412.09055","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hyperbolic regularizer and triplet loss are added to NU-MCC for single-view point cloud reconstruction, improving F1 on CO3D-v2.","lead":"This paper adds hyperbolic-space losses to a point cloud reconstructor and reports improved F1 scores on CO3D-v2. A generalist might read it because it offers a simple recipe for hierarchy-aware 3D reconstruction, but the method text contains major inconsistencies and omits the proposed hyperbolic Chamfer distance from the actual loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hyperbolic losses in Eqs. 10 and 12 are not well-defined as written; until clarified, the reported gains over NU-MCC cannot be attributed to hyperbolic geometry.","rationale":"The reader's weakest assumption identifies Eq. 10 and Eq. 12 as mathematically problematic, and my independent reading of the manuscript confirms this is the most load-bearing point. The central claim is an empirical comparison showing large gains over NU-MCC; those gains are supposed to come from hyperbolic regularization and triplet loss. But if those loss terms cannot be evaluated as written, the experimental numbers have no clear connection to the proposed hyperbolic mechanism. The paper gives no code, no error bars, and no detailed hyperparameter search, so there is no way to separate a genuine hyperbolic effect from an implementation correction or from normal training variance. I considered whether the more obvious issue is that HyperCD is not included in the total loss (Eq. 9), which is also true and further weakens the paper's claim about its contributions. However, the reader's identified concern is deeper: even the losses that are used are undefined, so the method itself is not reproducible. Because this concern directly undermines the strongest claim and the reader already recommended REJECT, no verdict change is needed.","tokens_in":12375,"tokens_out":3956,"duration_ms":43692,"concrete_test":"Request the authors' training code, or independently implement Eqs. 10 and 12 exactly as written using the stated k = -0.14 and run one forward-backward pass on the CO3D-v2 features. If the loss tensor is non-scalar, imaginary, or causes a backward error, the method as specified is not implementable. Then ask the authors to provide the corrected scalar definition; re-run the ablation baseline+LZ+LT versus NU-MCC with that corrected loss. If the corrected loss does not reproduce the reported 5.9% F1 improvement, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that HcPCR improves on NU-MCC by 5.9% in F1, 5.2% in completeness, and 5.6% in accuracy because of hyperbolic constraints. The only hyperbolic-specific terms in the total loss L = LN + LZ + LT (Eq. 9) are LZ and LT. Both are mathematically invalid as written. In Eq. 10, Gamma_P(W+) and Gamma_P(P+) are outputs of the Poincare-ball mapping and are therefore vectors; ordinary vector subtraction of these terms is not a meaningful operation in hyperbolic space, and max(0, vector) is undefined as a scalar loss. Even interpreted elementwise, the expression would not encode the stated goal that partial embeddings be closer to the center and whole embeddings closer to the edge, because no norm is taken. In Eq. 12, the loss feeds tangent-space vectors Psi_P(W+) and Psi_P(P+) into a hyperbolic distance d defined in Eq. 7 for Poincare-ball points; tangent-space vectors are not points in the ball, so the distance is not defined. Moreover, with the stated curvature k = -0.14, Eq. 7 contains sqrt(k) = sqrt(-0.14), which is imaginary unless an unstated convention or absolute value is introduced. As a result, the named hyperbolic losses cannot be computed in the form shown, and the reported performance improvements cannot be reproduced or verified from the manuscript alone. The paper does not state that a norm is intended or that a corrected scalar version of Eq. 10 is used, so the reader cannot determine whether the gains come from hyperbolic geometry or from an implicit, unstated modification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HcPCR, a single-view RGB-D point cloud reconstruction method built on NU-MCC, which embeds partial and complete point-cloud features into a Poincaré-ball model of hyperbolic space. The contributions are a hyperbolic Chamfer distance (Eq. 7), a regularization loss L_Z (Eq. 10) with adaptive margins, and a triplet loss L_T (Eq. 12), in addition to the NU-MCC loss L_N. The authors report improved F1, completeness, and accuracy over MCC and NU-MCC on CO3D-v2 and provide ablations over margins, curvature, and δ-hyperbolicity.","tokens_in":12729,"tokens_out":8627,"duration_ms":88008,"significance":"If the method worked as described, the paper would be a useful first application of hyperbolic geometry to single-view RGB-D point cloud reconstruction. The main comparison is against an external benchmark (F1, precision, recall), so the principal result is not circular, and the reported improvements are potentially interesting. However, the manuscript has serious mathematical gaps in the central loss definitions and the proposed HyperCD is not used in the total loss, so the current text does not support the claim that the gains come from hyperbolic constraints. The δ-hyperbolicity ablation (Table 2) is also weak evidence because it measures the effect of a regularizer explicitly designed to create a hierarchy.","major_comments":[{"comment":"L_Z = max(0, -Γ_P(W+) + Γ_P(P+) + γ/N) is a vector-valued expression: Γ_P maps to vectors in R^n, so max(0, vector) is undefined. Even interpreted elementwise, the expression does not encode the stated scalar margin between hyperbolic norms. A norm such as max(0, ||Γ_P(W+)|| - ||Γ_P(P+)|| + γ/N) is required. As written, this loss cannot be optimized and cannot be responsible for the reported gains.","section":"Hyperbolic Point Cloud loss, Eq. (10)"},{"comment":"L_T evaluates the hyperbolic distance d from Eq. (7) on Ψ_P(W+), Ψ_P(P+), and Ψ_P(P-). However, Ψ_P outputs tangent-space vectors in T_{z_P} P_n^k, whereas d is defined for points in the Poincaré ball through Möbius addition. The argument of d is therefore not in the domain of the distance. If the intended distance is the Euclidean norm of tangent vectors, then the loss is not a hyperbolic-distance triplet loss and the text must say so.","section":"Hyperbolic Point Cloud loss, Eq. (12)"},{"comment":"With the stated curvature k = -0.14, √k is imaginary and Eq. (7) is not real-valued. The limit in Eq. (8) is also inconsistent with Eq. (7): for the displayed formula one obtains approximately 2k||x-y|| as k→0, not 2||x-y||. The standard Poincaré distance uses 2/√|k| arctanh(√|k| ||x⊕y||). This error invalidates the distance metric that is central to the hyperbolic losses.","section":"Hyperbolic distance, Eqs. (7)-(8)"},{"comment":"The paper introduces a Hyperbolic Chamfer Distance in Eq. (7), but the total loss L = L_N + L_Z + L_T does not include it; L_N is described only as 'all losses in NU-MCC'. No experiment or ablation varies the use of HyperCD. Thus the central proposed distance is not actually used in the training objective, and the reader cannot tell whether the reported improvements are caused by hyperbolic geometry or by the extra L_Z and L_T terms.","section":"Method / Hyperbolic Point Cloud loss, Eq. (9)"},{"comment":"The quantitative comparisons report no error bars, number of runs, or statistical tests. The claimed average improvements of 5.9%, 5.2%, and 5.6% in F1 score, completeness, and accuracy could be within run-to-run variation. Given the invalid loss definitions in Eqs. (10) and (12), the experimental results are not reproducible from the manuscript as written.","section":"Table 1 / Results Comparison"}],"minor_comments":[{"comment":"The Poincaré ball is defined with radius 1/|k|, but the Möbius formulas in Eq. (3) and the distance in Eq. (7) correspond to the standard normalization with radius 1/√|k|; please reconcile the normalization conventions.","section":"Eq. (1)"},{"comment":"The prefactor '2p |k|λ_k(z_P)' in the tangent-space mapping appears to be a typographical error for a factor involving 2/√|k|; as printed, the expression is dimensionally inconsistent.","section":"Eq. (4)"},{"comment":"The text says 'δ is a hyperparameter that controls the separation degree between positive and negative samples', but Eq. (12) uses ε, not δ; this is a notation inconsistency.","section":"After Eq. (12)"},{"comment":"The δ-hyperbolicity values are reported without error bars or significance testing, and the improvement from 0.326 to 0.294 is expected because the regularizer is designed to push embeddings into a hierarchical arrangement; the claim of inherent hyperbolicity should be stated more cautiously.","section":"Table 2"},{"comment":"Since Lin et al. already proposed a Hyperbolic Chamfer Distance for point cloud completion, the paper should compare with that method or explicitly explain how the proposed distance differs beyond being used in a different reconstruction setting.","section":"Related work / experiments"}],"recommendation":"reject","confidential_remarks":"To the editor: the manuscript's central mathematical formulations are not well-defined, and the proposed HyperCD is absent from the training loss. These are load-bearing issues that preclude verification of the stated improvements. I recommend rejection, though a substantially rewritten version with corrected equations, a properly integrated HyperCD, and reproducible experimental details could be considered anew."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real hook—porting hierarchical losses into hyperbolic space for single-view point cloud reconstruction—but the method section as printed does not define the losses it uses. The reported F1 gains over NU-MCC are substantial, yet until Eqs. (7), (10), and (12) are corrected, the gains cannot be traced to hyperbolic geometry.\n\nWhat's new here: combining a hyperbolic regularizer and triplet loss with the NU-MCC backbone, plus an adaptive margin, and evaluating on CO3D-v2. That combination appears to be new. The ablation table is useful: adding the losses in Euclidean space barely moves F1, while doing the same in hyperbolic space jumps from 0.802 to 0.861. If the losses are intended to be what I suspect (a norm inside the max, distance evaluated after mapping back), the improvement is plausible and worth pursuing.\n\nSoft spots are concentrated in the method section. Eq. (7) uses sqrt(k) with k = -0.14, which is imaginary unless they mean sqrt(|k|). The regularization loss in Eq. (10) subtracts two Poincaré vectors and feeds the result to max(0, ·), which is undefined as a scalar; a norm is needed. The triplet loss in Eq. (12) passes tangent-space vectors to a distance defined for points in the ball. And the total loss in Eq. (9) never includes the promised HyperCD; it is introduced but not used. Any one of these would be a minor typo; all four together make the method unreproducible as written. The paper also reports no error bars and compares only against MCC and NU-MCC, which is thin.\n\nThe δ-hyperbolicity table is suggestive but mostly confirms that the regularizer pushes embeddings toward a tree-like layout; that is not an independent validation. The novelty is modest: HyperCD and the regularized triplet loss already exist in Lin et al. and Montanaro et al.; the contribution is the adaptation and the adaptive margin.\n\nI would not cite this in its present form, but I would not desk-reject it. The idea is sound and the empirical direction is promising. A serious referee should engage, send it back for major revision, and ask for corrected equations, released code, and error bars. My guess is that the errors are fixable typos, but I can't verify that without a revised manuscript.","headline":"Real idea, broken equations: the hyperbolic losses as printed cannot be computed, so the F1 gains over NU-MCC are not yet attributable to hyperbolic geometry.","tokens_in":13241,"tokens_out":2975,"would_cite":false,"duration_ms":30282,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that representing point cloud features in hyperbolic space, with a hyperbolic Chamfer distance and regularized triplet loss, improves single-view 3D reconstruction and reports average gains of about 6 percent in F1 score…","keywords":["hyperbolic space","point cloud reconstruction","RGB-D","Poincaré ball","Chamfer distance","triplet loss","hierarchical structure","CO3D-v2"],"falsifier":"Implement Eqs. (10) and (12) exactly as printed in a standard autodiff framework: if the losses do not produce a scalar (or training diverges) because the vector-valued max or the tangent-space-to-Poincaré distance is undefined, then the reported improvements cannot be attributed to hyperbolic geometry. Alternatively, retrain on CO3D-v2 with the curvature $k$ set to a value at which hyperbolic distance becomes Euclidean while keeping all other components unchanged; if the gains vanish, the hyperbolic metric is what carries the result.","tokens_in":12209,"feed_emoji":"📦","tokens_out":6164,"duration_ms":56008,"temperature":0.7,"pith_summary":"This paper tries to establish that moving feature embeddings for single-view 3D point cloud reconstruction from Euclidean space into hyperbolic space improves reconstruction quality. The reason to care is that object parts form tree-like hierarchies, and hyperbolic space can embed trees with much less distortion than Euclidean space. The proposed model, HcPCR, adds a hyperbolic Chamfer distance, a regularization loss that pushes partial clouds toward the center and complete clouds toward the boundary of a Poincaré ball, and a triplet loss that separates categories. On the CO3D-v2 benchmark, the authors report average gains over NU-MCC of 5.9% in F1 score, 5.2% in completeness, and 5.6% in accuracy.","feed_headline":"Hyperbolic losses lift 3D reconstruction F1 by 5.9%","feed_subtitle":"Embedding features in curved space captures part-whole structure, beating MCC and NU-MCC baselines.","key_machinery":"The central machinery is the Poincaré ball model of hyperbolic space, $\\mathbb{P}_k^n = \\{x \\in \\mathbb{R}^n : \\|x\\| < 1/|k|\\}$, together with two maps: a projection $\\Gamma_P$ from Euclidean vectors into the ball and a tangent-space map $\\Psi_P$ from the ball to a local Euclidean tangent plane. The paper defines a hyperbolic Chamfer distance by replacing the Euclidean point distance with the Poincaré distance $d(x,y) = \\frac{2}{\\sqrt{k}} \\operatorname{arctanh}(\\sqrt{k} \\| -x \\oplus_k y\\|)$, and adds two losses built on these maps: the regularization loss $L_Z = \\max(0, -\\Gamma_P(W^+) + \\Gamma_P(P^+) + \\gamma/N)$ with an adaptive margin computed by an MLP, and the triplet loss $L_T = \\max(0, d(\\Psi_P(W^+), \\Psi_P(P^+)) - d(\\Psi_P(W^+), \\Psi_P(P^-)) + \\varepsilon)$. Together they enforce that partial clouds sit closer to the center than whole clouds and that classes are separated by geodesic distance.","core_discovery":"The central claim is that hyperbolic geometry is the right setting for relating partial and complete point clouds. The paper's HcPCR model takes an RGB-D image, extracts global and local features with vision transformers, maps both into a Poincaré ball, and trains with a loss that combines the NU-MCC loss with a hyperbolic regularization term and a hyperbolic triplet term. The regularization term encodes a part-whole hierarchy, placing simpler parts near the center of the ball and larger parts nearer the boundary; the triplet term pulls together parts and wholes of the same class while pushing different classes apart. The authors argue that Euclidean space cannot represent the tree-like composition of objects without distortion, so the same losses that hurt in Euclidean space help once the geometry matches the data's hierarchical structure.","pith_inferences":["A natural extension the paper does not explore is treating the curvature $k$ as a learned parameter rather than a fixed value, which could remove a sensitive hyperparameter and adapt to datasets with different amounts of hierarchy.","The hyperbolic Chamfer distance could be combined with density-aware or other robust Chamfer variants, since the paper only tests the plain hyperbolic replacement and does not isolate the effect of outlier sensitivity.","The part-whole hierarchy losses might apply to multi-view reconstruction or neural fields, where partial observations are even more explicitly related to a complete object, though this is speculative and untested.","The paper does not report statistical significance or variance across runs; if the 5.9% F1 gain is within run-to-run noise, the hierarchy story would need stronger evidence."],"forward_implications":["The reported gains imply that existing transformer-based single-view reconstruction models can be improved by re-embedding their features in hyperbolic space and adding only two loss terms, without changing the backbone.","If the part-whole hierarchy claim holds, the same regularized triplet formulation should transfer to other tasks where partial observations relate to full structures, such as shape completion from occluded images.","The ablation results imply that regularization losses designed for hyperbolic space actively degrade Euclidean training, so applying these losses without the geometric change is counterproductive.","The curvature $k$ becomes a key hyperparameter; the sensitivity observed in ablations suggests that practitioners should tune $k$ per dataset or learn it during training.","The $\\delta$-hyperbolicity evaluation indicates that the learned feature space becomes more tree-like, offering a quantitative check on whether the model actually exploits hierarchy."],"supporting_citations":[{"why":"Defines the hyperbolic Chamfer distance that Eq. (7) adapts for reconstruction.","marker":"(Lin et al. 2023)"},{"why":"Introduces the hyperbolic regularization and variable margin that inspire the part-whole loss.","marker":"(Montanaro, Valsesia, and Magli 2022)"},{"why":"Provides the NU-MCC baseline architecture and loss $L_N$ on which HcPCR is built.","marker":"(Lionar et al. 2023)"},{"why":"Supplies the MCC baseline, the CO3D-v2 evaluation setup, and the transformer-based training approach.","marker":"(Wu et al. 2023)"},{"why":"Proves that trees embed in the hyperbolic plane with low distortion, motivating the use of hyperbolic space for hierarchical point clouds.","marker":"(Sarkar 2012a)"},{"why":"Supplies the $\\delta$-hyperbolicity measure used to show the embeddings become more tree-like.","marker":"(Khrulkov et al. 2020)"},{"why":"Introduces the Poincaré ball model and its embeddings, used for all hyperbolic mapping.","marker":"(Nickel and Kiela 2017)"},{"why":"Provides the CO3D-v2 dataset and its 51-category video collection used for training and evaluation.","marker":"(Reizenstein et al. 2021)"}],"fun_headline_variants":["Hyperbolic geometry sharpens single-view 3D reconstruction","Curved space lifts point cloud F1 by 5.9%","Hyperbolic Chamfer loss refines single-view point clouds","Poincare ball embedding boosts RGB-D reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training objective must be a well-defined, differentiable scalar function as written; in particular, the expression in the regularization loss subtracts and adds vector embeddings inside a max, and the triplet loss feeds tangent-space vectors into a hyperbolic distance, so the paper implicitly assumes some norm or valid mapping that it does not state.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic geometry sharpens single-view 3D reconstruction","Curved space lifts point cloud F1 by 5.9%","Hyperbolic Chamfer loss refines single-view point clouds","Poincare ball embedding boosts RGB-D reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2369,"prompt_tokens":864,"completion_tokens":1505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":1438}},"tokens_in":480,"tokens_out":1505,"duration_ms":11880,"temperature":1.0,"reasoning_tokens":1438,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:21:56.245940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement Eqs. (10) and (12) exactly as printed in a standard autodiff framework: if the losses do not produce a scalar (or training diverges) because the vector-valued max or the tangent-space-to-Poincaré distance is undefined, then the reported improvements cannot be attributed to hyperbolic geometry. Alternatively, retrain on CO3D-v2 with the curvature $k$ set to a value at which hyperbolic distance becomes Euclidean while keeping all other components unchanged; if the gains vanish, the hyperbolic metric is what carries the result.","supporting_citations":[{"cited_title":"D.; and Zhang, Z","cited_arxiv_id":null,"evidence_quote":"Defines the hyperbolic Chamfer distance that Eq. (7) adapts for reconstruction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the hyperbolic regularization and variable margin that inspire the part-whole loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NU-MCC baseline architecture and loss $L_N$ on which HcPCR is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the $\\delta$-hyperbolicity measure used to show the embeddings become more tree-like."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Poincaré ball model and its embeddings, used for all hyperbolic mapping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CO3D-v2 dataset and its 51-category video collection used for training and evaluation."}],"review_version":1}