{"id":"c401d9f1-8e39-48c0-ba5e-e880acccbc0c","arxiv_id":"2412.17951","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"HyperCD replaces the Euclidean distance in Chamfer Distance with arcosh(1 + alpha * squared distance), giving a gradient weighting that favors close point pairs and improves point cloud completion across multiple networks and benchmarks.","lead":"A new loss function for point cloud completion, Hyperbolic Chamfer Distance, applies an arcosh transformation to Euclidean point pair distances to weight close matches more heavily during training. Replacing the standard Chamfer Distance loss with it improves completion, reconstruction, and upsampling on several benchmarks, though gains are modest.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HyperCD's benefit may come from its tuned loss scale rather than the claimed distance weighting: arcosh(1+αr^2)≈√(2α)·r for small r, so α acts like a learning-rate multiplier, and the CD baselines are not given equivalent scale/lr tuning.","rationale":"Reader's weakest_assumption is essentially the same confound; I agree. The paper's Props 1–4 are mathematically correct, and the reported improvements are consistent, which is real evidence that something about the loss helps. But Prop 1 actually sharpens the concern: since matching is identical to Euclidean CD, any benefit must come from gradient reweighting or loss scale. Because the authors tune α but not the baseline's learning rate or loss scale, and because HyperCD≈√(2α)·L1-CD in the small-distance regime, the reported advantage may be an artifact of better-scaled optimization rather than the specific decreasing-weight mechanism. The absence of error bars and alpha values makes this hard to rule out. The test above would settle it, so the verdict stays conditional pending those controls.","tokens_in":20824,"tokens_out":10199,"duration_ms":96954,"concrete_test":"On ShapeNet-Part with CP-Net (and on ShapeNet-55 with SeedFormer), train the standard L1-CD baseline with loss multiplied by a scalar c, grid-searching c over the effective scale range spanned by HyperCD (c≈√(2α) over the α grid), and separately grid-search the learning rate for CD over the same multiplicative range, using identical seeds, epochs, and evaluation. If the best rescaled/lr-tuned CD reaches or beats HyperCD's reported CD/F1, the improvement is not attributable to the distance-dependent weighting. Also rerun HyperCD with 3–5 seeds and report mean±std and the selected α per dataset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1 (Eq. 8) shows that because g(r)=arcosh(1+αr^2) is strictly increasing, the nearest-neighbor matches under HyperCD are exactly the Euclidean matches; the method never changes correspondences, contrary to the Fig. 1 motivation. The only possible source of improvement is the per-pair gradient scale in Eq. 7: z(r)=2αr/√((1+αr^2)^2−1). But for small r (the typical regime for matched pairs in completion), z(r)≈√(2α), a constant, so HyperCD behaves like √(2α)·(L1-CD). The paper grid-searches α while explicitly keeping learning rate and all other hyperparameters at baseline values (Sec. 4), and reports no chosen α values and no error bars. Thus HyperCD is given an extra effective scale/learning-rate knob that the CD baselines are not. Table 7 includes other arcosh variants but not a rescaled L1-CD with tuned scale. The DCD comparisons are also implausible (Table 1: DCD+SeedFormer 24.52 vs CD 6.74; Table 7: DCD 5.74 vs L1-CD 4.16), undermining confidence in the loss-comparison setup. If a rescaled CD or a tuned learning rate closes the gap, the central causal claim fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hyperbolic Chamfer Distance (HyperCD), defined as arcosh(1 + α||x − y||^2) applied to each nearest-neighbor pair in the standard Chamfer Distance loss. The authors argue that this monotone reweighting gives larger gradients to close, well-matched point pairs and smaller gradients to distant, poorly matched pairs, which they call a position-aware mechanism. They report consistent improvements when HyperCD replaces CD in several point cloud completion networks on PCN, ShapeNet-55, and ShapeNet-34, and they also apply HyperCD to single-image reconstruction and point cloud upsampling. The paper includes propositions about the interchange of min and g, the monotonicity of the distance, and the limiting behavior of the gradient, plus an efficient algorithm whose cost is comparable to CD.","tokens_in":21154,"tokens_out":3274,"duration_ms":31763,"significance":"If the central claim is sustained, HyperCD is an attractively simple drop-in replacement for the Chamfer Distance loss: it requires one additional arcosh operation, preserves the nearest-neighbor matching step of CD, and is accompanied by public code and a small set of analytical propositions. The empirical gains are modest but appear consistent across multiple backbones and tasks, which makes the paper potentially useful to practitioners. However, the causal explanation offered for the gains is not established: Proposition 1 shows that the nearest-neighbor correspondences are identical to those of Euclidean CD, so the hyperbolic-space matching correction illustrated in Fig. 1 is not realized, and the reported experiments do not control for the extra loss-scale/learning-rate knob introduced by grid-searching α. Some reported baselines (notably DCD and KITTI Fidelity) are implausible, which weakens confidence in the loss-comparison setup.","major_comments":[{"comment":"The KITTI results contain impossible Fidelity values: the CD-ℓ2 Fidelity row reports 0.000 for both PoinTr and HyperCD + PoinTr, and hyperCD+FoldingNet drops from 7.467 to 2.214. A Fidelity of exactly 0.000 for a learned completion model on real LiDAR scans is not credible as a per-point accuracy measure and suggests a broken evaluation pipeline or a typo in the table. Please correct or explain these numbers; as printed, they vitiate the claim that HyperCD consistently improves performance on KITTI.","section":"Table 4, Sec. 4.1"},{"comment":"The DCD comparisons are implausible and undermine the loss-comparison setup. In Table 1, DCD + SeedFormer reports 24.52 average L1-CD versus 6.74 for CD + SeedFormer, a factor-of-3.6 degradation that is inconsistent with the published DCD results and with the moderate numbers reported elsewhere. In Table 7, DCD gives 5.74 on ShapeNet-Part while L1-CD gives 4.16, again contradicting the original DCD paper's claims. Please verify the DCD implementation and report the original authors' numbers or explain the discrepancy.","section":"Table 1, Table 7"},{"comment":"The central causal claim that HyperCD's gradient weighting improves completion is not separated from the effect of loss scale. For small matched-pair distances, which are the typical regime in completion, Eq. (7) gives z ≈ √(2α), so HyperCD behaves like √(2α) times the L1-CD gradient. The paper grid-searches α while explicitly keeping learning rates and all other hyperparameters at baseline values, so α serves as an extra effective learning-rate/scale knob that the CD baselines do not have; Fig. 5 confirms that α and lr jointly affect CD. To support the position-aware explanation, the authors should compare against a rescaled L1-CD (e.g., L1-CD with a tuned multiplicative scale) and against a CD trained with a learning rate tuned in an equally favorable manner.","section":"Eq. (7), Sec. 4 Implementation, Fig. 5"},{"comment":"Proposition 1 shows that because arcosh(1+αr²) is strictly increasing, the nearest-neighbor matches under HyperCD are exactly the Euclidean nearest-neighbor matches. Consequently, HyperCD never changes the correspondences, contradicting the Fig. 1 motivation that 'the mismatched point pairs in Euclidean space may be corrected' in hyperbolic space. The only difference from CD is the per-pair gradient weighting in Eq. (7). The paper should either reframe the contribution as a gradient-weighting scheme (dropping the matching-correction narrative) or provide evidence that the weighting mechanism, not scale, is what drives the gains, e.g., by comparing against another monotone transform with the same small-distance slope but a different large-distance profile.","section":"Proposition 1, Eq. (8), Fig. 1"},{"comment":"No error bars, confidence intervals, or multiple-seed results are reported. Several of the claimed improvements are small (e.g., Table 1: SeedFormer from 6.74 to 6.54; Table 2: SeedFormer CD-Avg from 0.92 to 0.86), and without run-to-run variance it is impossible to judge whether these differences are statistically meaningful. Please report mean and standard deviation over at least three seeds for the main completion tables, or otherwise justify the stability of the single-run numbers.","section":"Sec. 4, Tables 1-3"}],"minor_comments":[{"comment":"The phrase 'we also provide the provide experimental results beyond completion task' appears in both the abstract and the introduction; please remove the duplicated 'provide'.","section":"Abstract/Introduction"},{"comment":"The third bullet contains the typo 'HypperCD'; it should read 'HyperCD'.","section":"Sec. 1 Contributions"},{"comment":"The text says 'point cloud unsampling' in the evaluation-metrics paragraph; this should be 'upsampling'.","section":"Sec. 4, Upsampling"},{"comment":"The caption states that PU-GCN's superior performance is 'denoted as bold in the results', but no entries in the table are bold.","section":"Table 5 caption"},{"comment":"The paper performs a grid search over α but never reports the chosen α values for any experiment; please report the selected values for reproducibility.","section":"Sec. 4.2 Hyperparameters"},{"comment":"The reported per-iteration timings (0.4239±0.0019, 0.4298±0.0014, 0.5335±0.0368 seconds) would be more interpretable if the hardware and batch size were stated in the same paragraph.","section":"Sec. 4.2 Computation"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is simple and the code is promised, but the experimental validation currently contains several red flags: impossible KITTI Fidelity values, a clearly broken DCD baseline, and no control for the loss-scale confound. These are fixable within the manuscript's scope, so I do not recommend rejection, but the revisions are substantial. I would also ask the editor to ensure the authors address the discrepancy between the Fig. 1 motivation and Proposition 1 during revision, as this affects how the contribution is framed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: HyperCD is arcosh(1+αr²) applied to Chamfer distances. The paper is honest that the nearest-neighbor matches are exactly those of Euclidean CD (their Prop. 1), so the entire effect is per-pair gradient weighting. The math is correct and the implementation is trivial. The experiments show small but consistent improvements when HyperCD replaces CD across a range of networks and tasks, and source code is promised. That is the real contribution: a drop-in loss that sometimes helps.\n\nWhat's genuinely good: the analysis of the gradient weight in Prop. 3 and Fig. 3 is useful, and the computational overhead is negligible. The multi-task coverage (completion, SVR, upsampling) is broad. The paper doesn't hide the α grid search, though it never reports the chosen α values or error bars.\n\nWhere it gets soft. The stress-test note is on target. Because the function is strictly increasing, correspondences are identical to Euclidean CD. For small r, which is the regime of matched pairs in completion, the derivative approaches √(2α) — a constant. So α mostly acts as a global gradient scale, and with learning rate held fixed, HyperCD effectively gets an extra tuning knob that the CD baseline doesn't. The authors grid-search α and keep everything else fixed, so the reported gains could be from a better-scaled loss rather than from \"position-aware weighting.\" A rescaled L1-CD with tuned scale is the missing control, and it's absent from Table 7. Until that's done, the causal story is unverified.\n\nThere are also empirical red flags. Table 4 has 0.000 Fidelity for both PoinTr and HyperCD+PoinTr, which is suspicious. Table 1's DCD+SeedFormer at 24.52 is clearly a broken baseline, and Table 7's DCD at 5.74 vs L1-CD at 4.16 looks wrong given DCD's published behavior. These don't kill the paper, but they undermine trust in the loss-comparison setup.\n\nFinal word: worth a serious referee. The central idea is simple and plausible; the main claim is that the weighting helps, and that needs a proper control. If the authors add a rescaled baseline, report α and variance, and fix the DCD/KITTI issues, this could be a useful practical note. I'd send it to review, but with a request for those controls.","headline":"A one-line loss tweak with consistent but unproven gains: the hyperbolic framing is mostly decoration, and the scale confound is real, but it's worth refereeing.","tokens_in":21645,"tokens_out":2092,"would_cite":false,"duration_ms":19577,"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":"Replacing Chamfer Distance with a hyperbolic arcosh loss improves point cloud completion across benchmarks.","keywords":["point cloud completion","Chamfer distance","hyperbolic space","arcosh loss","density-aware Chamfer distance","point cloud upsampling","single-view reconstruction"],"falsifier":"Train the same completion backbone with the squared-Euclidean loss $d(r)=r^2$ and with HyperCD $\\operatorname{arcosh}(1+\\alpha r^2)$ while matching the initial gradient magnitude; if the squared-CD model matches HyperCD's benchmark error, then the arcosh nonlinearity and its specific weighting curve are not the operative cause. A second check is to sweep $\\alpha$ widely on a held-out set to see whether the reported results depend on a narrow tuned range.","tokens_in":20646,"feed_emoji":"🧊","tokens_out":5713,"duration_ms":47334,"temperature":0.7,"pith_summary":"The paper proposes Hyperbolic Chamfer Distance (HyperCD), a drop-in replacement for the Chamfer Distance loss in point cloud completion. HyperCD computes the usual nearest-neighbor Chamfer sum but passes each squared Euclidean distance through $\\operatorname{arcosh}(1+\\alpha r^2)$, with $\\alpha>0$. The claim is that this simple change reweights gradients so that already-close point pairs are preserved while poor matches are corrected gradually, and that models trained with it outperform the same models trained with CD or DCD. On the PCN, ShapeNet-55, and ShapeNet-34 benchmarks the paper reports state-of-the-art completion results, plus gains in single-image reconstruction and point cloud upsampling. The authors argue this is the first use of hyperbolic geometry in point cloud completion.","feed_headline":"Arc-cosh loss beats Chamfer distance on point cloud completion","feed_subtitle":"Replacing the standard point-cloud loss with a hyperbolic version lowers error and smooths surfaces on three benchmarks.","key_machinery":"The central object is the HyperCD distance $d(x_{ij},y_{ik})=\\operatorname{arcosh}(1+\\alpha\\|x_{ij}-y_{ik}\\|^2)$, applied inside the Chamfer min-sum over nearest neighbors. Its load-bearing property is the gradient weight $z_{ij}=2\\alpha r/\\sqrt{(1+\\alpha r^2)^2-1}$, which Proposition 4 shows is strictly decreasing in the Euclidean match distance $r$; this is what redirects optimization toward preserving good matches. Proposition 1 justifies computing the minimum in Euclidean space first, keeping the algorithm almost as cheap as ordinary Chamfer Distance.","core_discovery":"HyperCD replaces the per-pair distance $d(x_{ij}, y_{ik})$ inside Chamfer Distance with $\\operatorname{arcosh}(1+\\alpha\\|x_{ij}-y_{ik}\\|^2)$, $\\alpha>0$. Because $\\operatorname{arcosh}(1+\\alpha r^2)$ is strictly increasing in $r$, the nearest-neighbor match found in Euclidean space is unchanged, so the extra cost is just one $\\operatorname{arcosh}$ per match. During backpropagation the gradient carries a factor $z_{ij}=2\\alpha r/\\sqrt{(1+\\alpha r^2)^2-1}$ that decreases as the matched distance $r$ grows; the paper's Proposition 4 proves this monotone decrease. The intended effect is that accurate matches receive larger gradient weight and are stabilized, while distant outlier matches receive smaller weight and are adjusted incrementally. Empirically the paper shows this weighting yields lower Chamfer error and smoother surfaces than CD and DCD across several backbones, and reports state-of-the-art numbers on PCN, ShapeNet-55, and ShapeNet-34.","pith_inferences":["Because the paper compares HyperCD against CD and DCD but not against other monotone transforms (e.g., squared Euclidean CD or a logarithmic loss) with matched gradient scales, the specific arcosh form may not be essential; a reweighting family with similar close-pair emphasis could reproduce the gains.","A testable extension is to use the same position-aware weighting as a plug-in for other set-matching losses (e.g., in 2D/3D registration or shape retrieval), where the same outlier-versus-fine-detail trade-off occurs.","The reported dependence of optimal $\\alpha$ on learning rate suggests that loss-scale calibration, rather than geometry, may drive part of the improvement; ablating with per-backbone learning-rate rescaling would separate the two effects."],"forward_implications":["Training any of the tested completion backbones (PCN, FoldingNet, PMP-Net, PoinTr, SnowflakeNet, PointAttN, SeedFormer) with HyperCD in place of CD lowers per-point Chamfer error on PCN.","On ShapeNet-55 and ShapeNet-34, HyperCD with SeedFormer improves both CD and F-Score@1% over the CD-trained baseline, including on unseen categories.","HyperCD helps downstream generative tasks: swapping CD for HyperCD in SnowflakeNet improves single-view reconstruction, and swapping it into four upsampling networks improves CD, Hausdorff, and point-to-surface metrics on PU1K.","HyperCD's computational cost is close to that of CD: the only added operation per match is one arcosh, in contrast to the heavier full hyperbolic distance."],"supporting_citations":[{"why":"Introduces Density-aware Chamfer Distance, the main alternative loss whose gradient behavior HyperCD is designed to improve upon.","marker":"[25]"},{"why":"Supplies the PCN dataset, the PCN baseline network, and the standard completion benchmark protocol used for evaluation.","marker":"[13]"},{"why":"Provides the PoinTr backbone and the ShapeNet-55/34 benchmark setup with difficulty levels and evaluation metrics.","marker":"[19]"},{"why":"Supplies the SeedFormer backbone that yields HyperCD's headline results on ShapeNet-55 and ShapeNet-34.","marker":"[11]"},{"why":"Motivates the use of hyperbolic embeddings for point clouds by showing they capture compositional, tree-like structure.","marker":"[26]"},{"why":"Supplies the PointAttN backbone used in the PCN comparison, including the DCD-trained baseline that HyperCD outperforms.","marker":"[1]"}],"fun_headline_variants":["HyperCD: arcosh-weighted Chamfer distance wins on point cloud tasks","Hyperbolic Chamfer Distance: better point cloud completion via arcosh","HyperCD: Hyperbolic Chamfer distance smooths surfaces and beats SOTA","New loss HyperCD: Hyperbolic Chamfer for robust point cloud completion","HyperCD: A hyperbolic reweighting of Chamfer distance for point clouds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that HyperCD's distance-dependent gradient weighting, not incidental differences like loss scale or the grid-searched $\\alpha$, causes the reported improvements; the paper does not compare against other monotone transforms or rescaled CD baselines to rule those out.","fun_headline_variants_meta":{"raw":{"variants":["HyperCD: arcosh-weighted Chamfer distance wins on point cloud tasks","Hyperbolic Chamfer Distance: better point cloud completion via arcosh","HyperCD: Hyperbolic Chamfer distance smooths surfaces and beats SOTA","New loss HyperCD: Hyperbolic Chamfer for robust point cloud completion","HyperCD: A hyperbolic reweighting of Chamfer distance for point clouds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001198,"raw_usage":{"total_tokens":4980,"prompt_tokens":1029,"completion_tokens":3951,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":3852}},"tokens_in":645,"tokens_out":3951,"duration_ms":26312,"temperature":1.0,"reasoning_tokens":3852,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:32.424085+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same completion backbone with the squared-Euclidean loss $d(r)=r^2$ and with HyperCD $\\operatorname{arcosh}(1+\\alpha r^2)$ while matching the initial gradient magnitude; if the squared-CD model matches HyperCD's benchmark error, then the arcosh nonlinearity and its specific weighting curve are not the operative cause. A second check is to sweep $\\alpha$ widely on a held-out set to see whether the reported results depend on a narrow tuned range.","supporting_citations":[{"cited_title":"Density- aware chamfer distance as a comprehensive metric for point cloud completion,","cited_arxiv_id":null,"evidence_quote":"Introduces Density-aware Chamfer Distance, the main alternative loss whose gradient behavior HyperCD is designed to improve upon."},{"cited_title":"Pointr: Diverse point cloud completion with geometry-aware transformers,","cited_arxiv_id":null,"evidence_quote":"Provides the PoinTr backbone and the ShapeNet-55/34 benchmark setup with difficulty levels and evaluation metrics."},{"cited_title":"Rethinking the compositionality of point clouds through regularization in the hyperbolic space,","cited_arxiv_id":null,"evidence_quote":"Motivates the use of hyperbolic embeddings for point clouds by showing they capture compositional, tree-like structure."}],"review_version":1}