acceptodds
Under review as a conference paper at ICLR 2027

Sub-Riemannian structure for oriented point clouds

Abstract

Many frameworks dealing with 3D (oriented) point clouds define locality in the ambient Euclidean space: -nearest-neighbor graphs and radius balls in are critical to define which points are "close to one another" and should exchange information. This can unintentionally merge sampled surfaces that are close in space but geometrically distinct such as the two faces of a thin leaf. Then, features are more prone to leak across boundaries which hinder performance. To address this, we propose to define a sub-Riemannian contact metric on position-orientation space : points are lifted to with a surface normal, and "motion" is constrained to a non-holonomic horizontal distribution in which spatial velocity must stay in the plane orthogonal to the normal which can reorient at a cost. From this, we derive an explicit score, which locally approximate the distance, to compare oriented points for downstream tasks. We report improved segmentation and correspondence accuracy near contact when using the score as a drop-in change inside PointNet++ aggregation mechanism as well as reduced leakage when using it to simulate heat diffusion on oriented points.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
shape analysis, geometry processing, sub-riemannian contact geometry, point clouds

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UZrVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a principled geometric prior for defining locality in oriented point clouds by lifting samples to position–orientation space M = R^3 × S^2 and endowing it with a contact sub-Riemannian (SR) structure in which spatial motion is constrained to be tangential (orthogonal to the local normal) while reorientation incurs a weighted cost. Leveraging the nilpotentization at each point (Mitchell’s theorem), the authors derive an explicit, SE(3)-invariant pairwise semi-metric (“contact score”) based on the 5D Heisenberg group that exhibits linear growth with normal offset, thereby discouraging “short-circuits” across thin structures. They demonstrate benefits as a drop-in neighborhood mechanism in PointNet++ for segmentation (Pheno4D) and correspondence (FAUST), and as a kernel for diffusion on oriented point sets using a Sinkhorn-normalized operator, reducing leakage across contacting surfaces.

Critical strengths

1.The metric prior addresses a pervasive failure mode of Euclidean neighborhoods in thin, layered, or contacting structures—relevant across 3D learning and geometry processing.

2.Demonstrations suggest the proposed contact score can be a drop-in component in common pipelines to reduce leakage and improve accuracy without architectural overhauls.

Critical weaknesses

1.The paper omits both the AI Use Statement and the Reproducibility Statement.

2.Figure 2 (Page 2): The top text labels suffer from severe horizontal truncation—specifically, the upper portions of the words "Slide" (in "Slide in reoriented tangent plane") and "Tilt" (in "Tilt back to match normals", which is even misread as "Liit" in the text layer) are cut off, exhibiting typical traits of AI-generated figure text distortion or careless image cropping.

3.The contact score is a semi-metric (no triangle inequality), which may affect graph construction properties; implications are not deeply analyzed.

4.Choice of the weight w is fixed and crucial for balancing spatial vs. orientation scales; practical guidance, adaptivity, or learning of w is not fully explored.

5.Solving for ϕ in H(a^2, τ) requires a root-finding step; numerical stability, efficiency, and differentiability in batched GPU settings are not thoroughly discussed.

Questions for authors

Please refer to weakness.

1 comment

Response to Reviewer UZrV

Thank you for the review, we hope the following answers will address your concerns.

About AI use and reproducibility statements

We added these in the revised version of the paper.

About Figure 2

Good catch for the cropped figure, we corrected it in the revised version. As written in our new AI use statement, we did not use LLM assistance for any of the figure, including this one, which has been entirely done by one of the authors.

The score is a semi-metric

You are right that the triangle inequality can fail. It fails only at scales that are large relative to ww but none of our uses requires it. Proposition 4.1 gives ℓw=dCC+o(ϵ)\ell_w = d_{CC} + o(\epsilon), and dCCd_{CC} is a true metric. Within a neighbourhood of size ϵ\epsilon, the triangle defect is small at only o(ϵ)o(\epsilon)

  • kNN grouping only ranks candidates for each query, so it needs no metric property. Farthest-point sampling and the decoder's interpolation remain Euclidean.
  • Sinkhorn kernels need a symmetric, positive kernel, which we have. The kernel does not need to be SPD, so the balanced operator can have negative eigenvalues. This is why Eq. (12) uses a generator and not fractional powers of the operator.
  • The separation results (Theorem B.2, Corollary B.3) rely on the bounds of Lemma B.1, not on the triangle inequality.
  • Fast neighbour search cannot use triangle-inequality pruning, as in metric trees, but an exact radius bound replaces it. If ℓw2≤L\ell_w^2 \le L, Lemma B.1 gives ρ2≤L\rho^2 \le L and ∣ζ∣≤L/(2πw)\lvert\zeta\rvert \le L/(2\pi w), hence ∥r∥2≤L+L2/(4π2w2)\lVert r\rVert^2 \le L + L^2/(4\pi^2 w^2).

On the choice of w

ww is a length: turning the normal by an angle α\alpha costs the same as sliding by wαw\alpha. It sets the distance at which a difference in orientation matters as much as a difference in position. Write ε\varepsilon the distance to the kk-th same-surface neighbour, δ\delta the gap to separate, σn\sigma_n the normal noise, in radians and RR the curvature radius. Here are practical guidance:

  1. Lower limit, from separation: w>ε2/(2πδ)w > \varepsilon^2/(2\pi\delta) suffices on planes (Corollary B.3). The gap must cost more than the farthest same-surface neighbour.
  2. Upper limit, from noise: noise makes same-surface neighbours look offset by about εσn\varepsilon\sigma_n, and a larger ww charges that offset too. On planes, the best ww is about 0.3 ε/σn0.3\,\varepsilon/\sigma_n (about 3.4ε3.4\varepsilon at 5°).
  3. Upper limit, from curvature: along a direction of curvature κ\kappa, same-surface costs grow by a factor 1+w2κ21 + w^2\kappa^2. On spheres, separation degrades between w≈1.3Rw \approx 1.3R and 5.2R5.2R.

From our experiments on real data:

  • Learning: separation stops improving at about w=2hw = 2h to 4h4h. We recommend choosing ww there (Reviewer Hyvg, Q3).
  • Diffusion: the best ww is about σ\sigma on curved shapes but at least 8σ8\sigma on stacked sheets (Reviewer Hyvg, Weakness 4), curvature plays a role here and as you pointed out, learning could help.

On the root-finding: stability, efficiency and differentiability

There is no iterative solver per pair. We invert one fixed monotone function, in batch on the GPU. The score needs the turning angle ϕ\phi of the optimal arc. ϕ\phi solves f(ϕ)=sf(\phi) = s, with f(ϕ)=(ϕ−sin⁡ϕ)/(1−cos⁡ϕ)f(\phi) = (\phi - \sin\phi)/(1 - \cos\phi) and s=4∣τ∣/a2s = 4\lvert\tau\rvert/a^2.

  • ff is monotone on (0,2π)(0, 2\pi), so we tabulate it once and invert it by binary search with linear interpolation.
  • Near 2π2\pi, ff is evaluated through u=2π−ϕu = 2\pi - \phi to avoid cancellation.
  • The edge cases τ=0\tau = 0 and a≈0a \approx 0 return a2a^2 and 4π∣τ∣4\pi\lvert\tau\rvert exactly.

Efficiency. On the GPU, a training step at 4,096 points is 20–22% slower than with Euclidean grouping (Reviewer Hyvg, Q1) so it is still pretty fast to compute.

Differentiability. In our experiments no gradient flows through HH. If one is wanted, there is no need to differentiate through the inversion, because the partial derivatives are explicit in ϕ\phi:

∂H∂∣τ∣=2ϕ,∂H∂a2=ϕ2cot⁡ϕ2.\frac{\partial H}{\partial \lvert\tau\rvert} = 2\phi, \qquad \frac{\partial H}{\partial a^2} = \frac{\phi}{2}\cot\frac{\phi}{2}.

The second derivative diverges as a→0a \to 0 with τ≠0\tau \neq 0. There, HH is only Lipschitz in aa. Gradient-based training through the score is untested.

Reviewer_HyvgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a normal-aware neighborhood construction method for point clouds to reduce cross-surface feature mixing in thin, folded, and near-contact regions. It replaces Euclidean kNN with a distance-based ranking in position–normal space using a Heisenberg-inspired sub-Riemannian metric. The method is evaluated on point cloud segmentation, dense correspondence, and heat diffusion, without refining the input normals.

Critical strengths

Euclidean kNN can connect nearby points from different surfaces. The proposed score uses positions and normals to select neighbors and makes cross-layer connections more costly. Its geometric design directly addresses the problem. Beyond synthetic examples, the paper evaluates plant segmentation, human-body correspondence, and heat diffusion. Improvements in some contact regions provide evidence of practical value.

Critical weaknesses
  1. The evidence is too limited to establish generality. Evaluation relies on one cylinder configuration, three diffusion shapes, and a single learning backbone. These results do not establish consistent gains across gaps, curvatures, or sampling conditions, particularly without a complete and consistent set of baselines.

  2. The central connectivity claim lacks sufficient direct validation. Improved segmentation and correspondence do not establish fewer cross-layer connections. Normal agreement is explicitly favored by the score and is therefore not independent evidence of correct connectivity; stem/leaf classification likewise cannot detect erroneous connections between distinct leaves. Evaluation against ground-truth connectivity is needed.

  3. The score conflates layer separation with within-surface smoothness. It penalizes legitimate curvature-induced normal changes as well as cross-layer connections, potentially rejecting valid neighbors or introducing directional bias. The paper does not establish whether improved layer separation comes at the expense of same-surface neighborhood coverage, especially in high- or varying-curvature regions.

  4. The diffusion comparison does not isolate the proposed kernel's contribution. CONTACT and the Gaussian baseline use different evolution schemes, leaving kernel, bandwidth, and temporal evolution effects confounded. Lower final errors therefore cannot be attributed to the grouping rule alone. The acknowledged early-time mismatch with mesh diffusion also leaves unresolved whether reduced leakage preserves accurate within-surface propagation.

Questions for authors
  1. How sensitive are accuracy, cross-layer edges, and runtime to point count, grouping k, PCA neighborhood size, normal noise, and w?

  2. Could you test more layer gaps, folds, curvatures, and sampling conditions, include failure cases, and consistently compare LIN and QUAD with clearly stated parameter choices?

  3. When w is increased to separate layers, are valid neighbors in high-curvature directions lost? How does the method handle point clouds sampled from non-manifold geometry, such as intersecting sheets or branching junctions? Given the ambiguity of normals at these junctions, can the grouping preserve valid connections rather than incorrectly treating them as cross-layer edges?

  4. Does increasing w sacrifice valid neighbors in high-curvature regions? Is the local approximation still reliable at coarse network levels?

1 comment

Response to Reviewer Hyvg

Thank you for this detailed review, we ran the experiment you asked for and hope they addressed your concerns

On Sensitivity to point count, grouping k, PCA neighbourhood size, normal noise and w

We ran more parameter analysis reported in the revised paper. Here is a breakdown:

  • Cross-surface edges (shortcuts) are reduced in every setting. The reduction is smaller with larger PCA neighbourhoods, denser sampling and, on FAUST, larger kk which is expected.
  • Accuracy: the FAUST gain holds in most change of setting. The Pheno4D gain does not hold when kk changes too much.
  • Runtime: +20–22% per training step compared to euclidean neighborhood, whatever kk is and the overhead grows with point count.
  • Noise sets a floor on the gap that can be separated, and too large a ww makes it worse.

And here is some of the analysis:

The training setting is k=32k = 32, PCA normals from 24 neighbours (Pheno4D) or 16 (FAUST), 4,096 points and w=4hw = 4h. We changed one factor at a time.

SettingPheno4D mIoU ↑ (EUC / CONTACT)FAUST contact-stratum error ↓ (EUC / CONTACT)
Training setting0.808 / 0.8370.077 / 0.049
k=16k = 160.762 / 0.6710.089 / 0.089
k=64k = 640.827 / 0.8160.069 / 0.048
PCA from 12 / 8 neighbours0.801 / 0.7960.093 / 0.070
PCA from 48 / 32 neighbours0.841 / 0.8230.099 / 0.064
2,048 points0.735 / 0.8280.047 / 0.027
8,192 points0.738 / 0.7510.134 / 0.123

FAUST: CONTACT is ahead in most of the setting Pheno4D: two settings stand out: a clear loss at k=16k = 16 and a clear gain at 2,048 points. The other differences are within seed variance. Larger PCA neighbourhoods average the normals across the gap, so less of the gap shows up in ζ\zeta, and the reduction weakens.

Runtime. CONTACT makes a training step 20–22% slower for kk from 16 to 64 so it is still pretty fast.

Failure cases reported:

  • Too large a ww on curved surfaces: on spheres (Q1), and in diffusion (reported in Appendix D).
  • Too small a ww: on the stacked sheets in diffusion, leakage is 67% at w=σw = \sigma, 27% at 4σ4\sigma and 0.09% at 8σ8\sigma.
  • Touching surfaces: with no gap there is no separation

Folds are untested, but a fold combines two cases we did test: the stacked sheets and highly curved regions.

Does a larger w lose valid neighbours in curved regions?

Yes. The loss comes from the angular term, not from ζ\zeta, and beyond w≈2hw \approx 2h–4h4h it buys no further separation. The two parts of the score react differently to curvature:

  • ζ\zeta does not see curvature. The second-order terms of its two projections cancel, so ζ=O(∥r∥3)\zeta = O(\lVert r\rVert^3) on a smooth surface, and ζ=0\zeta = 0 exactly on spheres and cylinders.
  • The angular term does. Along a direction of normal curvature κ\kappa, the normals turn by α≈κr\alpha \approx \kappa r, so a2≈r2(1+w2κ2).a^2 \approx r^2(1 + w^2\kappa^2).

We added a measure over all queries on the Pheno4D test scans and FAUST test shapes, at 4,096 points in Appendix D.

Non-manifold geometry

Non-manifold geometry is not handled by design. The score assumes one normal per point, so it cannot tell a valid edge across a junction from a cross-layer edge. Intersecting sheets are untested, and so are PCA normals at junctions, where the normal itself is unreliable, we added this as limitation in the revised paper.

Is the approximation reliable at coarse network levels?

We make no claim there. Proposition 4.1 is asymptotic: it holds for fixed ww as the two points get close. The spacing grows at coarse levels, so they leave that regime

Diffusion: kernel and evolution rule are confounded

We ran both kernels under both evolution rules: the continuous generator and fixed steps. Positions, normals, bandwidth, masses and source are the same in every run. We report results in Appendix D. We observe the following:

The kernel effect: under either rule, CONTACT is more accurate at the final time on all three shapes, and at every time on the sheets. Early times on the curved shapes: the Gaussian kernel is more accurate, through at least 8τ8\tau on the bunny and until about 4τ4\tau–7τ7\tau on the dragon. We no longer claim accurate early-time propagation on curved shapes. The early spike is mostly due to the evolution rule The continuous generator keeps at least a fraction e−t/τe^{-t/\tau} of the heat at the source, whatever the kernel. The Gaussian kernel spikes just as much under it. The remaining gap under fixed steps is a difference between the kernels.

Section 5.3 now reports the comparison under a common rule (fixed steps), with the full table in Appendix D.

Reviewer_u3SdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a contact sub-Riemannian construction for defining locality in point clouds equipped with surface normals. Each sample is represented in position-orientation space R3×S2\mathbb{R}^{3}\times S^{2}, where admissible spatial motion is restricted to the plane orthogonal to the current normal, while normal reorientation incurs a weighted cost. A pairwise score derived from the five-dimensional Heisenberg tangent model locally approximates the associated Carnot-Carath'eodory distance; an extension to unoriented normals minimizes the score over normal sign choices. The score is used for neighborhood grouping in PointNet++ and for constructing affinities for Sinkhorn-normalized diffusion operators. The reported experiments cover neighborhood-based reconstruction of two coaxial cylinders, binary stem-leaf segmentation on Pheno4D, template correspondence on registered FAUST meshes, and heat diffusion on the Stanford bunny, dragon, and stacked sheets.

Critical strengths
  1. The method translates tangential-motion constraints and normal reorientation costs into a computable, rigid-motion-invariant locality score, with an extension to unoriented normals. The same score is used for neighborhood grouping and diffusion affinity construction.

  2. The learning comparisons keep the PointNet++ backbone and normal inputs fixed within each setting. Pheno4D and FAUST are split by plant and subject identity, respectively, with means and standard deviations reported over three seeds.

  3. The FAUST experiments separately evaluate contact regions, examining the geometric situation motivating the method. With PCA normals, CONTACT achieves a mean contact-region error of 0.0568, compared with 0.0828 for EUC and 0.0668 for LIN (Table 3).

Critical weaknesses
  • The related-work comparison overstates the distinction from prior work by contrasting graph construction with PDE solving: Bekkers et al.\ (2018a) already perform sub-Riemannian graph grouping using nilpotent distance approximations. The paper-specific contribution should be delineated around the surface-normal motion constraint, symmetric pair coordinates, and resulting H5H^5-based score.

  • The Pheno4D evaluation does not directly measure the motivating leakage between nearby leaves. All leaf instances share one target class, and no cross-instance neighborhood statistic is reported. A missing diagnostic is the proportion of cross-leaf neighbor connections in a common set of leaf-contact regions, using the existing instance annotations.

  • The unqualified linear-growth interpretation after Eq.(9) omits the required limiting conditions. For fixed a>0a>0, Eqs.~(8)-(9) give quadratic variation near zero normal offset; the pure-normal relation ℓw2=4πw∣ζ∣\ell_w^2=4\pi w|\zeta| concerns a different regime. The derivation should specify these regimes and distinguish the score from its square.

  • The ablation lacks a LIN control matched to CONTACT's pure-normal asymptotic coefficient, such as a2+4πw∣ζ∣a^2+4\pi w|\zeta|, while holding the angular term and normal-sign handling fixed. LIN and QUAD are also missing from the cylinder and diffusion comparisons.

Questions for authors
  • Relative to \textit{Nilpotent Approximations of Sub-Riemannian Distances for Fast Perceptual Grouping of Blood Vessels in 2D and 3D} by Bekkers, Chen, and Portegies (2018), which components of the proposed surface-normal motion constraint, symmetric pair coordinates, and H5H^5-based score are inherited, adapted, or newly derived?

  • Using the existing leaf-instance annotations, can you report cross-leaf neighbor connection rates for CONTACT, EUC, LIN, and QUAD? Please evaluate a common set of queries near other leaves, specify the proximity criterion, and keep the point samples and neighborhood size identical across methods.

  • Under which limiting conditions does the claimed linear dependence on normal offset hold, and does it refer to ℓw\ell_w or ℓw2\ell_w^2? Please distinguish the small-offset expansion for fixed a>0a>0 from the pure-normal relation ℓw2=4πw∣ζ∣\ell_w^2=4\pi w|\zeta|, and qualify the corresponding statements in the contribution list and Section 4.

  • Can you add a coefficient-matched LIN control, a2+4πw∣ζ∣a^2+4\pi w|\zeta|, while keeping ww, the angular term, and normal-sign handling fixed? Please also include LIN and QUAD in the cylinder and diffusion experiments under the same evaluation protocols as CONTACT.

  • Section 5.2 interprets ∣ζ∣/∥r∥|\zeta|/\lVert r\rVert as deviation from the query's tangent plane, although the two projections in ζ=12r⋅(nP+nQ)\zeta=\frac{1}{2}r\cdot(n_P+n_Q) can cancel. Could you clarify this interpretation and check whether the reported comparisons of neighborhood flatness hold when using the actual normalized query-plane deviation, ∣r⋅nP∣/∥r∥|r\cdot n_P|/\lVert r\rVert?

1 comment

Response to Reviewer u3Sd

Thank you for the precise questions, we ran the experiment you asked for and hope they addressed your concerns


What is from Bekkers et al. (2018)?

We inherit three things: the nilpotent approximation used as a fast grouping distance, the state space, and sign minimisation. The motion constraint is adapted. The pair coordinates, the use of the exact distance as the score, its theory, and its use inside networks are new. Our sentence "graph construction rather than a PDE solver" did not distinguish the two works, since Bekkers et al. also build a graph so we rewrote this part (see the revised related work section in blue).


Cross-leaf connection rates

Concretely, scores that charge ζ\zeta remove a large portion of cross-leaf connections near contact. For this specific measure, CONTACT ties its tuned linear and quadratic approximations LIN, QUAD. Of course, keeping cross leaf points does not implies a loss in accuracy for the trained model.

  • Data: all 36 Pheno4D test scans with same protocol as in Table 2.
EUCPTPQUADLIN 2πLIN 4πCONTACT
Cross-leaf rate0.2810.1730.1690.1540.1490.158
Share of EUC10.820.650.650.630.64

Closer leaves are harder to separate and the cross-leaf rate falls as the sampling gets denser, because "within hh" then means physically closer leaves:


When does growth in the normal offset become linear?

The statement is about ℓw2\ell_w^2, not ℓw\ell_w, and it is exact only when the pair has no tangential or angular displacement (a=0a = 0). For a typical neighbour (a>0a > 0, small ζ\zeta) the penalty is quadratic. What separation actually uses is a lower bound that is linear in ∣ζ∣\lvert\zeta\rvert for every aa. We correct the text accordingly.

Below, s=4w∣ζ∣/a2s = 4w\lvert\zeta\rvert/a^2 measures the normal offset relative to the tangential displacement.

Regimeℓw2\ell_w^2Growth in ∣ζ∣\lvert\zeta\rvert
a>0a > 0, small ζ\zetaa2+12 w2ζ2/a2+…a^2 + 12\,w^2\zeta^2/a^2 + \dotsquadratic
all a,ζa, \zetamax⁡{a2,2πw∣ζ∣}≤ℓw2≤a2+4πw∣ζ∣\max\{a^2, 2\pi w\lvert\zeta\rvert\} \le \ell_w^2 \le a^2 + 4\pi w\lvert\zeta\rvertbetween linear bounds

The separation still works because it only needs the lower bound ℓw2≥2πw∣ζ∣\ell_w^2 \ge 2\pi w\lvert\zeta\rvert, which holds for every aa:

  • a neighbour on another sheet, at gap δ\delta, pays at least 2πwδ2\pi w\delta;
  • a neighbour on the same sheet (ζ≈0\zeta \approx 0) pays about its squared distance ε2\varepsilon^2.

Matched LIN, and LIN and QUAD in the cylinder and diffusion experiments

We added control over LIN-4π= a2+4πw∣ζ∣a^2 + 4\pi w\lvert\zeta\rvert because it is the exact upper envelope of ℓw2\ell_w^2 We trained LIN-4π with three seeds, keeping ww, the angular term and the sign handling unchanged. The other rows are as printed in Tables 2–3.

GroupingPheno4D mIoU ↑Pheno4D acc. ↑FAUST PCA error ↓FAUST PCA contact ↓FAUST oriented error ↓FAUST oriented contact ↓
EUC0.750 ± 0.0270.913 ± 0.0140.0137 ± 0.00100.0828 ± 0.01100.0098 ± 0.00050.0434 ± 0.0085
LIN-4π (new)0.797 ± 0.0230.936 ± 0.0090.0125 ± 0.00020.0625 ± 0.00910.0098 ± 0.00080.0383 ± 0.0096
CONTACT0.835 ± 0.0230.950 ± 0.0070.0121 ± 0.00070.0568 ± 0.00730.0095 ± 0.00040.0385 ± 0.0050

Cylinder. Same setting as Sec. 5.1:

EUCPRODPTPQUAD 4πLIN 2πLIN 4πCONTACT
First neighbour on the correct cylinder0%0%0%0%0%100%100%
Cross-cylinder edges242412121200

LIN-4π succeeds as well, as expected.

Diffusion. Each score was swapped into the kernel of Section 5.3, with everything else unchanged. Values are the discrepancy from mesh heat at the final time (%, lower is better):

KernelSheetsBunnyDragon
CONTACT0.302.433.92
LIN 2π0.295.408.73
LIN 4π0.2810.1213.68
QUAD 4π50.003.817.63
PROD (angle only)50.003.103.45
PTP66.679.1515.48
Gaussian (same evolution rule)66.677.5610.45

We report the following:

  • Sheets: every kernel that charges ζ\zeta at least linearly separates them. QUAD, PROD, PTP and the Gaussian kernel leak.
  • Curved shapes: linear penalties are much worse than CONTACT, and the larger the coefficient, the worse they are:
    • on a curved, sampled surface, same-surface pairs have small but non-zero ζ\zeta;
    • a linear charge penalises these pairs at full slope;
    • ℓw2\ell_w^2 adds only 12 w2ζ2/a212\,w^2\zeta^2/a^2 for them.

Is ∣ζ∣/∥r∥\lvert\zeta\rvert/\lVert r\rVert a deviation from the query's tangent plane?

The two projections in ζ\zeta can cancel and ζ=0\zeta = 0 on any sphere or cylinder. So ∣ζ∣/∥r∥\lvert\zeta\rvert/\lVert r\rVert is a normalised symmetric offset.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.