acceptodds
Under review as a conference paper at ICLR 2027

Manifold-Aware Persistent Homology Representation for Point Cloud Registration with Topology Preservation

Abstract

Point clouds sampled from real-world objects often exhibit complex non-convex geometry, in which global structure such as connectivity, holes and part adjacency is essential for shape understanding, yet topological analysis based on Euclidean distances measures proximity in ambient space rather than intrinsic connectivity and can create shortcuts across narrow gaps or concave boundaries. We introduce a manifold-aware topological representation that computes persistent homology from geodesic distances to provide faithful global shape cues, and condition a registration network on it: TopoReg routes the case-level descriptor into a standard coarse-to-fine backbone at a single deep level, so that the gain comes from the representation rather than from added capacity. We further prove that any stable descriptor built on the Euclidean filtration changes by at most an amount proportional to the gap width when a narrow gap is bridged, whereas the graph geodesic metric separates the two sides by an unbounded amount. Experiments on the large-scale lung CT dataset Lung250M-4B and the KITTI autonomous driving dataset show state-of-the-art performance and robust generalization.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Point cloud registration, Rigid and non-rigid alignment, State-space model

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_u1PPOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TopoReg for point cloud registration. It computes persistent homology using graph geodesic distances and builds a six-dimensional descriptor from persistence lifetimes. This descriptor is added to a PointPWC-based network through input features and FiLM conditioning. The method also uses a local edge-length consistency loss. The authors provide theoretical analysis and evaluate the method on Lung250M-4B and KITTI.

Critical strengths

The descriptor is compact and can be added to an existing registration network. The KITTI results show a clear improvement over the methods listed in the paper. The mean Chamfer distance decreases from 0.2206 m for PointMamba, the strongest listed baseline, to 0.1944 m for TopoReg. This suggests that the full system may have useful empirical benefits.

Critical weaknesses

The main idea and motivation of this paper are very close to “TopoPointPWC: Manifold Topology-Aware Point Cloud Registration via Persistent Homology,” ICLR 2026 GRaM Workshop, Tiny Paper Track. That work already uses persistent homology on local neighborhood graphs to guide hierarchical point cloud registration. Moreover, Figure 2 of the current submission still labels the proposed method “TopoPointPWC.” Prior work has explored persistent homology for 3D shape registration and point cloud alignment. Please clarify how the proposed use of persistent homology differs from these earlier approaches.

On Lung250M-4B, the mean TRE decreases from 2.21 mm for PointPWC to 2.18 mm for TopoReg. This is an absolute improvement of only 0.03 mm, or about 1.4%. The median improves by only 0.02 mm. Seems effectiveness of author's method is limited.

Questions for authors

Please address the concern raised in the Weaknesses section.

3 comments
AuthorsOpenReview ↗

Response to Reviewer u1PP [1/3]

Dear Reviewer u1PP,

Thank you for the detailed and precise review. Your three weaknesses identified the gaps that mattered most: the relationship to the earlier workshop report was not stated, the experiments did not isolate the descriptor from the auxiliary loss, and the proof of Theorem 4.2 was incorrect as written. We address all three below.


Weakness 1: Overlap with the workshop paper, and incremental contribution

Weakness 1. "The paper has substantial overlap with an existing workshop paper, and the additional contribution appears incremental."

We thank the reviewer for laying out the specific overlaps; the relationship should have been stated and was not.

Question 1. "What is the relationship between TopoReg and the earlier TopoPointPWC paper? Please explain the identical table entries, deformation statistics, and method name in Figure 2. Which contributions and results are new in this submission?"

We should have stated this relationship and did not. The revision adds an explicit statement, and we answer the four specific points below.

On the identical Table 4 entries. The 24 numerical entries, together with the median displacement of 11.91 mm and median smoothness of 1.76, are baseline-row statistics computed once on the same public split with the same preprocessing. They are shared because they describe the dataset and the baselines, not our method. Every row in Table 4 that reports our model is newly computed for this submission. The revision marks reused baseline rows explicitly so that the provenance of each row is unambiguous.

On the Figure 2 label. In an earlier visualisation we compared against an early topological variant; when that panel was dropped, the name of our own method was deleted by mistake and the older label was left in its place. The figure content has always been ours. The revision restores the correct label, "TopoReg".

On what is new. The differences from the earlier version are at the component level:

ComponentWorkshop versionThis submission
Diagram representationpersistence image6-D H0H_0/H1H_1 lifetime statistics
Conditioningtopology gating (per-point)single-layer pure-multiplicative FiLM at L4
Auxiliary losscurriculum-weightedlocal edge-length consistency, λ=0.05\lambda = 0.05 (Lung) / 0.010.01 (KITTI)
Parameter cost—+34,048 (+0.44%) over the PointPWC backbone

On what the submission contributes. The earlier version is a non-archival workshop report and is therefore not a prior publication of these results. The filtration construction itself is established theory, and we now cite that lineage. What this submission contributes is the cross-domain transfer of that construction into a registration architecture, where it acquires a function beyond description: the descriptor modulates the feature pathway through FiLM and measurably changes what the network can fit. That transfer, together with the controlled measurement of what the filtration choice costs, is what the earlier version does not report.

AuthorsOpenReview ↗

Response to Reviewer u1PP [2/3]


Weakness 2: Experiments do not isolate the components

Weakness 2. "The experiments do not isolate the benefits of the new components or sufficiently support the topology-preservation and SOTA claims."

We agree this was the central evidential gap, and we thank the reviewer for naming it so precisely.

Question 3. "Can the authors provide the missing PointPWC-plus-L_topo control and a direct comparison with the earlier TopoPointPWC formulation? Can they also report uncertainty estimates for the small improvement on Lung250M-4B?"

The control you asked for. We ran it: PointPWC with the same edge-length loss, descriptor and FiLM removed. Its parameter count is 7,720,236, identical to the PointPWC baseline, which confirms no topological signal enters the network, while the loss stays active.

ConfigurationDescriptorLtopo\mathcal{L}_{\text{topo}}Three runs (Lung TRE, mm)Mean ± std
PointPWC + Ltopo\mathcal{L}_{\text{topo}} (control)noyes2.31 / 2.22 / 2.282.270 ± 0.037
TopoReg (full)yesyes2.18 / 2.20 / 2.182.187 ± 0.009
Difference0.083 mm

The comparison against the earlier formulation is given by the component-level table in our response to Weakness 1, which specifies exactly which elements differ.

Both configurations contain the auxiliary loss, so the comparison isolates the descriptor-plus-FiLM branch. The gap is about 2.2× the control's spread and about 9× that of the full model.

The uncertainty you asked for. Three training seeds plus a three-by-three sampling matrix:

Training seedSubsample 1Subsample 2Subsample 3
02.202.202.20
12.272.232.23
22.232.202.22

Resampling within a fixed training seed changes TRE by at most 0.04 mm, so subsampling is not the dominant variance source; the training seed contributes more. Both spreads are now reported so that the mean differences can be read against them.


Weakness 3: Theoretical issues — shortcut edges and the Theorem 4.2 proof

Weakness 3. "There are also problems with the theoretical explanation. If a shortcut edge already exists in the Euclidean-weighted neighborhood graph, computing shortest paths cannot remove it... In addition, the proof of Theorem 4.2/A.2 assumes that a path whose length equals the Euclidean distance must contain only one edge. This is incorrect: several edges along the same straight line can give the same total distance."

We thank the reviewer for both objections: the first concerns how our claim is worded, the second identified a genuine error in the proof.

Question 2. "How does the implemented graph construction prevent or remove cross-boundary shortcut edges? How should the statement or assumptions of Theorem 4.2 be corrected?"

On shortcut edges: you are right that an existing edge cannot be removed. It keeps its Euclidean length, so its endpoints retain that graph distance. Our claim concerns penalty rather than removal: the geodesic metric forces detours around boundary-crossing pairs. The magnitude of that penalty is measurable, and we report it here:

Quantity (27 lung cases, 2048-point subsample, k=20k=20, fixed seed)Value
Median ratio dG/dEd_G / d_E1.206 ± 0.102
95th percentile of dG/dEd_G / d_E1.830 ± 0.585
Share of pairs with dG/dE>1.5d_G/d_E > 1.514.6 ± 16.9%
Same share at k=6k=6 / k=30k=3060.3% / 13.2%

Roughly 15% of sampled pairs pay a detour of more than 50%, while the rest are essentially unaffected. This is what "penalty, not removal" means quantitatively, and it also explains why the effect weakens as kk grows: larger neighbourhoods create more shortcuts, so the detectable penalty shrinks.

It also bears on the theory: the strict inequality of Theorem 4.2 requires an additional hypothesis, but the mechanism it describes does not depend on that hypothesis holding everywhere. On real data the penalty is present and sizeable for a clear minority of pairs, which is what the method relies on.

On the proof: you are correct that the hypothesis was missing. The claim that an equal-length path contains only one edge is false — collinear unit edges give the same total length. The corrected statement is L≥wL \ge w, with equality exactly when some pair realising the gap is joined by graph edges that follow the straight segment in order. The appendix proof is rewritten as a triangle-inequality argument, and the condition is stated inside the theorem.

We would add that the correction concerns the boundary case, not the operative regime: the inequality L≥wL \ge w still guarantees that graph distance never understates the separation, and the measured distribution above shows that it exceeds it substantially for the pairs that matter.

AuthorsOpenReview ↗

Response to Reviewer u1PP [3/3]


On the topology-preservation claim

Weakness 2 (continued). "The graph-level analyses also do not isolate whether the proposed descriptor reduces topology errors in the predicted deformation... Better input-graph statistics and lower registration error are not sufficient evidence that the new components preserve the topology of the output deformation."

We appreciate this distinction; it is one the original draft did not make.

This is a fair criticism. We added an output-side measurement: for each test case we compute the same 6-D persistence descriptor on the predicted warped cloud and on the ground-truth target cloud, and take their relative difference.

The median relative difference is about 20% (p90 ≈ 0.85). The metric is also uncorrelated with registration accuracy — the variant with the lowest descriptor difference (a learned readout) has the worst TRE. We therefore do not claim that the method preserves the topology of the output deformation. The revision states this in the limitations: the method improves registration accuracy, and output-level persistence consistency is left as future work.

We thank the reviewer again for the time spent on this paper, and in particular for the two observations that changed it most: that the equal-length-path hypothesis in Theorem 4.2 was missing, and that the control experiment was needed to make the attribution meaningful. We hope the revisions above address the concerns raised.

Reviewer_CyosOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to use a geodesic-distance driven persistent homology features to augment the typical deep features used to register two point sets. This is known as scene flow estimation and can also be used in piecewise-rigid or deformable registration. A classical cost-volume scene flow estimator is augmented through FiLM and input feature encoding.

Critical strengths

The pipeline is sensible and I believe it would be reproducible - although the gains from adapting such a complicated mechanism would be marginal.

Critical weaknesses
  • The paper claims the 'manifold-aware PH' to be a new thing. However, using PH with different distances, including the graph-geodesic distances have been done repeatedly in the literature.
  • I also don't understand why we compute 9D features and not pass the entire PH into the network. Also why don't we learn some topological features via PersLay [https://arxiv.org/abs/1904.09378] instead? This type of approaches are now being adopted more and more. See for example CumPerLay [https://arxiv.org/abs/2510.12795].
  • Why is Topformer omitted from the comparisons?
  • Paper's theorems are not new, e.g. stability. Not sure why these are stated in the paper.
  • Thm. 4.2 is also not new and it is one of the foundations of manifold learning - which the paper completely misses out. In fact I even teach and prove that theorem in my class. That whole section is a bit over-stated.
  • Coherent point drift, Carriere et al., Chen et al. etc are all cited twice.
  • Please cite thin-plate spline interpolation.
  • Theoretical analysis begins by 'Our theory establishes an asymmetry between the Euclidean filtration and the graph geodesic metric'. What do you mean by asymmetry here?
  • Eq. 15: what is s here?
  • How can we interpret Figure 2? It is really hard to see through different colors and the colormap is not specified. Can't you visualize different point sets in different colors to aid visual quality?
  • Persistent augmented networks can also be interesting here: https://arxiv.org/pdf/2604.08469v1. Also see: https://arxiv.org/pdf/2602.14228, and https://arxiv.org/pdf/2311.03087v3 (for different choices of distances). And [https://arxiv.org/pdf/2012.07621v3] for the shortest path version.
  • There are also many other scene flow datasets (in self-driving) where the paper can show evaluations on.
Questions for authors

Please see weaknesses. Though, my concerns are hard to fix in a minor revision.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Cyos [1/3]

Dear Reviewer Cyos,

Thank you for the detailed list of weaknesses. We address every one below. Two of them are addressed by clarifying the scope of our claim and by adding a controlled ablation that separates the filtration choice from the descriptor readout. The remaining items are clarifications, corrections, and added citations.


Weakness 1: Novelty of "manifold-aware PH"

Weakness 1. "The paper claims the 'manifold-aware PH' to be a new thing. However, using PH with different distances, including the graph-geodesic distances have been done repeatedly in the literature."

We thank the reviewer for this observation, which helps us state the scope precisely.

The filtration construction is established, and we cite the line of work you point to, including the density-based metric learning and spectral-method formulations. Our claim is not about the construction but about what is done with it.

What we propose is the cross-domain transfer of that theory into a registration architecture, where the descriptor acquires a role it does not have in the descriptive literature: rather than being reported as an output, it becomes a conditioning signal (via FiLM) that modulates the feature pathway and measurably changes what the network can fit. That transfer, and the accompanying measurement of what the filtration choice costs, is the new attempt; the construction itself is the tool.

Filtration (same backbone, same seed)KITTI ChamferLung TRE
Graph-geodesic (ours)0.1945 m2.18 mm
Euclidean0.2200 m2.22 mm
Error increase when replaced+13.2%+1.8%

We do not claim that graph persistence is new; we claim that quantifying this cost end-to-end is, and that the descriptor-plus-FiLM configuration is a new practice in representation learning for registration.


Weakness 2: Why 9-D features rather than the full diagram, and why not a learned readout

Weakness 2. "I also don't understand why we compute 9D features and not pass the entire PH into the network. Also why don't we learn some topological features via PersLay instead? This type of approaches are now being adopted more and more. See for example CumPerLay."

We appreciate this question, which goes to the design choice rather than a presentation detail.

Three reasons, and we ran the comparison rather than asserting it.

(i) Cost. A learned readout processes the diagram at every forward pass. Our six statistics are computed once per cloud and cached offline, so inference incurs no additional forward cost and the module adds only 34,048 parameters (+0.44%).

(ii) Stability. FiLM is multiplicative and sensitive to the scale of its conditioning input; a bounded, normalised statistic is markedly more stable to train.

(iii) Measurement. We implemented three alternative readouts on the same backbone and training budget:

ReadoutLung TREOutcome
6-D lifetime statistics (ours)2.18 mmreference
32-D rich lifetime statistics2.23 mmno gain from added capacity
PersLay-style learned aggregation (top-16 bars, weighted DeepSet)3.44 mmoverfits (97 training cases)
Descriptor only, auxiliary loss removed2.34 mmworse

The learned readout was the worst of the four. We state this as an empirical finding under our data regime, not as a general claim that PersLay is inferior — it is a more expressive family, and we now discuss it and CumPerLay explicitly in the related work.

We also note that the readout and the filtration are separate axes: the ablation above varies the readout; the table in Weakness 1 varies the filtration. Both are now reported.

AuthorsOpenReview ↗

Response to Reviewer Cyos [2/3]


Weakness 3: TopFormer omitted from the comparisons

Weakness 3. "Why is Topformer omitted from the comparisons?"

We thank the reviewer for pointing us to TopFormer, which is indeed the closest topology-aware work.

TopFormer (Fung et al., CVM 2024) does not use persistent homology. It couples surface geodesic distance with normal-based direction into a pairwise structural encoding consumed by Transformer attention. The mathematical object, the level at which topology acts, the task, and the metrics all differ:

TopFormer (CVM'24)TopoReg (ours)
Topological objectgeodesic distance + normal, per point pairpersistent homology (H0H_0/H1H_1 lifetimes)
Where it actspairwise encoding to attention6-D descriptor to FiLM conditioning
Taskrigid registration (estimate a transform)non-rigid per-frame registration
MetricsIR / FMR / RR / RRE / RTElandmark TRE / Chamfer
BackboneKPConv + coarse-to-fine + RANSACPointPWC scene-flow pyramid

Its official implementation is not public, so a same-pipeline reproduction is not possible and a cross-pipeline number would be uninformative. We did, however, run a mechanism-level control: we implemented TopFormer's pairwise structural encoding (geodesic distance + normal, queried sparse-through-dense) inside our backbone.

Encoding (same backbone)Lung TRE
6-D lifetime descriptor (ours)2.18 mm
TopFormer-style geodesic + normal (8-D)2.21 mm

The gap is small and we attribute it mainly to its additional use of normal-direction information. We state that this is a single-seed control.


Weaknesses 4 and 5: Theorems and Section 4 are not new

Weakness 4. "Paper's theorems are not new, e.g. stability. Not sure why these are stated in the paper."

We accept this characterisation and thank the reviewer for stating it plainly.

Weakness 5. "Thm. 4.2 is also not new and it is one of the foundations of manifold learning - which the paper completely misses out. In fact I even teach and prove that theorem in my class. That whole section is a bit over-stated."

We appreciate the reviewer pointing to the manifold-learning foundation, which the earlier draft did not cite.

Agreed on both counts, and the framing has been changed. Standard persistence stability is not our contribution; Theorem 4.1 is an application of it to the six statistics, and we now present it as such rather than as a result. Likewise, the relation between graph distance and the gap width is standard in manifold learning; we have added that lineage to the related work.

The reason we retained a theoretical section at all is narrower: to justify which of the six statistics can be trusted under what conditions (the mean and entropy bounds need equal-count and positive-lifetime assumptions), and to make the H1H_1 behaviour explicit. Section 4 has been rewritten as a short justification section rather than as a list of new theorems, and the "new" framing has been removed throughout.


Weaknesses 6 and 7: Duplicate citations and thin-plate spline

Weakness 6. "Coherent point drift, Carriere et al., Chen et al. etc are all cited twice."

Thank you for catching these; the duplicate entries have been merged.

Weakness 7. "Please cite thin-plate spline interpolation."

Thank you for the pointer; we have added the citation.

Both fixed. The duplicate entries (CPD, Carrière et al., Chen et al.) have been merged into single bibliography entries. Thin-plate spline interpolation is now cited where the displacement-field parameterisation is discussed, and we have corrected the interpolation wording.


Weakness 8: What "asymmetry" means

Weakness 8. "Theoretical analysis begins by 'Our theory establishes an asymmetry between the Euclidean filtration and the graph geodesic metric'. What do you mean by asymmetry here?"

We thank the reviewer for flagging this; the term was poorly chosen.

The term was poorly chosen. What we meant is that the two sides bound different objects:

SideQuantity boundedBound
Euclideana descriptor change (Rips stability)C0wC_0 w
Grapha pairwise-distance change between the two point setsL−wL - w

The original sentence invited the reading that both bounds refer to the same quantity. The revision names the two objects explicitly and drops the word "asymmetry".

AuthorsOpenReview ↗

Response to Reviewer Cyos [3/3]


Weakness 9: What is ss in Eq. 15

Weakness 9. "Eq. 15: what is s here?"

Thank you for spotting the undefined symbol.

s<ws < w bounds the spacing between consecutive inserted bridge points in the bridged construction. When the bridged graph contains the full ordered chain including its two endpoint edges, that chain has length ww, so the minimum graph distance between the original sets decreases by at least L−wL - w. We had left ss undefined; it is now named in the text and listed in the symbol table.


Weakness 10: Figure 2 readability

Weakness 10. "How can we interpret Figure 2? It is really hard to see through different colors and the colormap is not specified. Can't you visualize different point sets in different colors to aid visual quality?"

We appreciate these concrete suggestions for improving the figure.

We have revised the figure: the two point sets are drawn in distinct colours with a labelled colour scale. The figure is a qualitative illustration: non-adjacent pairs may be connected by longer graph paths, while an existing edge keeps its Euclidean length.

We also note what happened to the method label in Figure 2. In an earlier visualisation we compared against an early topological variant; when that panel was dropped, the name of our own method was deleted by mistake and the older label was left in its place. The figure content has always been ours. The revision restores the correct label, "TopoReg".


Weaknesses 11 and 12: Additional references and datasets

Weakness 11. "Persistent augmented networks can also be interesting here: [four references]."

We thank the reviewer for these references, which we have added.

Weakness 12. "There are also many other scene flow datasets (in self-driving) where the paper can show evaluations on."

We appreciate this suggestion and have addressed the scope explicitly.

The four persistent-augmentation references have been added to the related work, and the shortest-path formulation you cite is now discussed alongside the density-based and spectral-method variants.

On additional scene-flow datasets: we agree this would strengthen the empirical case, but we do not claim results on datasets we have not run. We have added it to the limitations as an explicit scope statement, and the KITTI evaluation we do report is per-frame Chamfer on the odometry split, which we describe as driving-domain evidence rather than as a scene-flow benchmark result.

We thank the reviewer again for the time spent on this paper, and in particular for the two observations that changed it most: that the equal-length-path hypothesis in Theorem 4.2 was missing, and that the control experiment was needed to make the attribution meaningful. We hope the revisions above address the concerns raised.

Reviewer_a57ROpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to add topological data descriptors to help the registration of point clouds. To this end, the topological data descriptors are computed via standard filtration using geodesic distance on a sampled graph of the point cloulds, instead of the extrinsic Euclidean distance. Then, they compute local statistical descriptors that are added to the spatial representation of the points. These features are then used in a classical registration pipeline such as PointPWC-Net which predics a non-rigid displacement for every source point. In particular, the baseline is really PointPWC method. Section 4 of the paper provides a theoretical analysis that proves stability results for their method. The authors show some results on Lungs data as well as on the KITTI database. The improvements over the baseline PointPWC particularly visible on the KITTI database.

Critical strengths

The paper tends to show a result that could be expected and rather important: adding topological features helps the registrattion of point clouds.

Critical weaknesses
  • The principal weakness in my opinion is the lack of empirical evidence for robustness of their results: in view of the small gain on the lungs dataset, a multiple seeds test is necessary. The authors only report 1 seed, a sort of confidence interval must be reported.

  • I have trouble understanding theorem 4.2 which seems to be incorrect in its current formulation. See the questions for authors section. Maybe an hypothesis is missing.

Questions for authors

Multiple seeds results must be shown. How do the results depend on the sampling of the point clouds ? it seems to me that this information is not reported. Are you sure that theorem 4.2 is correct: can it be applied to P ={(x,0,0); i=0,\ldots, 40} A = 0 and B = (40,0,0). Then, the 20-NN graph should be connected and maximum edge length is 20. However, d_G(A,B) =40 ???

2 comments
AuthorsOpenReview ↗

Response to Reviewer a57R [1/2]

Dear Reviewer a57R,

Thank you for your careful reading. Your two weaknesses point to the two places where the paper was genuinely under-supported: the robustness of the small Lung gain, and the formulation of Theorem 4.2. We address both below, with the additional runs and the corrected statement.

We are also grateful for the counterexample in your questions, which pinpoints the equality case we had conflated.


Weakness 1: Robustness of the small Lung gain

Weakness 1. "The principal weakness in my opinion is the lack of empirical evidence for robustness of their results: in view of the small gain on the lungs dataset, a multiple seeds test is necessary. The authors only report 1 seed, a sort of confidence interval must be reported."

We thank the reviewer for raising this directly: the robustness of the small Lung gain was indeed not established, and it is a fair objection.

We agree this was a real gap, and we have run the multi-seed experiment. Two sources of variance matter here: the training seed, and the topology subsample. We report both.

Across training seeds (350-point topology subsampling; the control is PointPWC with the same edge-length loss but without the descriptor and FiLM):

ConfigurationRun 1Run 2Run 3Mean ± std
PointPWC + Ltopo\mathcal{L}_{\text{topo}} (control)2.312.222.282.270 ± 0.037
TopoReg (full)2.182.202.182.187 ± 0.009
Difference0.083 mm

The gap of 0.083 mm is about 2.2× the control's standard deviation (0.037) and about 9× that of the full model (0.009). We therefore no longer describe the gain as being of unclear stability, but we also do not overstate it: the improvement over the unmodified PointPWC baseline in Table 1 is 0.03 mm, and it is the comparison against the same-loss control that carries the attribution.

Across point samples (three subsampling seeds within each of three training seeds, separate Lung configuration):

Training seedSubsample 1Subsample 2Subsample 3Max change
02.202.202.200.00
12.272.232.230.04
22.232.202.220.03

Resampling within a fixed training seed changes TRE by at most 0.04 mm, so the topology subsample is not the dominant variance source. The revision reports both tables so that the mean differences can be read against their spreads.


Weakness 2: Correctness and hypothesis of Theorem 4.2

Weakness 2. "I have trouble understanding theorem 4.2 which seems to be incorrect in its current formulation. See the questions for authors section. Maybe an hypothesis is missing."

We thank the reviewer for this careful reading. The formulation was indeed wrong in the case identified, and we are grateful for the counterexample.

Question 1. "Are you sure that theorem 4.2 is correct: can it be applied to P = {(x,0,0); i = 0, …, 40}, A = 0 and B = (40,0,0). Then, the 20-NN graph should be connected and maximum edge length is 20. However, dG(A,B)=40d_G(A,B) = 40???"

You are right, and the hypothesis that was missing is exactly the one your example identifies. In it the 20-NNNN graph is connected, its longest edge has length 20 (below the gap width 40), and the unit edges form a path of length 40, so L=w=40L = w = 40. The original text asserted a strict detour where only a non-strict one holds.

The corrected statement is L≥wL \ge w, with equality exactly when some pair realising the gap is joined by graph edges that follow the straight segment in order. The proof is a triangle-inequality argument: each path from a∈Aa \in A to b∈Bb \in B has length at least ∥a−b∥2≥w\lVert a - b \rVert_2 \ge w; equality requires both that the pair realise the gap and that every step points from aa toward bb. A strict detour therefore occurs iff every gap-realising pair lacks such an ordered chain, and the excess L−wL - w is the penalty the graph metric imposes.

The revision states the condition inside the theorem rather than in the surrounding prose, and the appendix proof is rewritten accordingly.

Why the two filtrations still separate: the H1H_1 loop

The same example explains why our method uses H1H_1 rather than the H0H_0 merge scale. For H0H_0, a merge occurs at the smallest possible largest-edge-weight along a connecting path, which in your example is already 20 under both metrics. For H1H_1, however, total path length determines when the loop fills.

Take the four corners of a rectangle with sides 0<a≤b0 < a \le b. The 2-NN graph is a four-cycle. The loop appears at scale bb under both metrics, but fills at

a2+b2(Euclidean)versusa+b(graph).\sqrt{a^2 + b^2} \quad \text{(Euclidean)} \qquad \text{versus} \qquad a + b \quad \text{(graph)}.

The maximum finite H1H_1 lifetime — one of the six inputs of TopoReg — records this difference directly, which is why the descriptor is not merely a re-encoding of Euclidean geometry.

AuthorsOpenReview ↗

Response to Reviewer a57R [2/2]


Question 2: Dependence on the sampling of the point clouds

Question 2. "How do the results depend on the sampling of the point clouds? It seems to me that this information is not reported."

We thank the reviewer for asking this; the sampling dependence had not been reported, and we now report it.

The three-by-three matrix reported above provides the quantitative answer: with the training seed fixed, changing the topology subsampling seed moves TRE by at most 0.04 mm across all nine cells (range 2.20–2.27 mm). We had not reported this previously; it is now in the revision.

We thank the reviewer again for the time spent on this paper, and in particular for the two observations that changed it most: that the equal-length-path hypothesis in Theorem 4.2 was missing, and that the control experiment was needed to make the attribution meaningful. We hope the revisions above address the concerns raised.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers [2/2]

3. Additional controls: readout, encoding, and filtration

Three further controls vary one axis at a time on the same backbone and training budget. Entries are Lung TRE (mm) and KITTI Chamfer (m).

Axis variedVariantResult
Readout6-D lifetime statistics (ours)2.18
Readout32-D rich lifetime statistics2.23
ReadoutPersLay-style learned aggregation3.44
Readoutdescriptor only, auxiliary loss removed2.34
EncodingTopFormer-style geodesic + normal (8-D)2.21
Filtrationgraph-geodesic (ours), KITTI / Lung0.1945 / 2.18
FiltrationEuclidean, KITTI / Lung0.2200 / 2.22

Readout. The learned readout was the worst of the four; it overfits at 97 training cases. The 32-D statistics gave no gain over the 6-D summary, which indicates that capacity is not the limiting factor. These are single-seed controls, and the exact gaps will be established with multi-seed runs.

Encoding. TopFormer does not use persistent homology; it couples geodesic distance with normal direction into a pairwise structural encoding. We implemented that encoding inside our backbone. The small gap is plausibly attributable to its additional use of normal information (8-D versus 6-D).

Filtration. Replacing the geodesic filtration by a Euclidean one raises the KITTI Chamfer error by 13.2% and the Lung TRE error by 1.8%, with every KITTI percentile worsening (median 0.1877 → 0.2081; 75th 0.2229 → 0.2585; max 0.3813 → 0.4799).

The filtration construction itself is established theory, which we now cite. What the paper proposes is its transfer into a registration architecture, where the descriptor acquires a role it does not have in the descriptive literature: rather than being reported as an output, it conditions the feature pathway through FiLM. That transfer, and the measurement above of what the filtration choice costs, is the new attempt; the construction is the tool.

4. Uncertainty and sampling

Two sources of variance matter: the training seed and the topology subsample. Both are now reported. The table gives TRE (mm) for three subsampling seeds within each of three training seeds, in a separate Lung configuration.

Training seedSubsample 1Subsample 2Subsample 3Max change
02.202.202.200.00
12.272.232.230.04
22.232.202.220.03

Resampling within a fixed training seed changes TRE by at most 0.04 mm, so the topology subsample is not the dominant variance source; the training seed contributes more. Together with the three-run spreads in Section 2, this gives the basis for reading the mean differences.

AuthorsOpenReview ↗

Response to All Reviewers [1/2]

Dear Reviewers,

We thank all three reviewers for the care they brought to this paper. Below we summarise the work shared across the individual responses: the corrections that apply to the paper as a whole, the control experiments that were missing, and the resulting changes to the manuscript. The three point-by-point responses contain the detailed answers.

1. Shared corrections

Theorem 4.2 and its hypothesis. The statement was too strong. The general relation is L≥wL \ge w, with equality exactly when some pair realising the gap is joined by graph edges that follow its straight segment in order. The collinear-points example in a57R's review is exactly the equality case. The appendix proof has been rewritten as a triangle-inequality argument. This concerns the boundary case rather than the operative regime: L≥wL \ge w still guarantees that graph distance never understates the separation, and on real data the excess is sizeable for the pairs that matter (Section 3).

Shortcut edges: penalty, not removal. We agree that an existing edge cannot be removed by computing shortest paths. The claim in the revision is that the geodesic metric penalises boundary-crossing pairs by forcing detours. The magnitude is measured rather than asserted, and it is reported in Section 3.

Relationship to the earlier non-archival report. The revision states this relationship explicitly. The 24 numerical entries in Table 4, together with the median displacement and smoothness values, are baseline-row statistics computed on the same public split; rows that report our model are newly computed, and the revision marks the provenance of each row. The earlier report is non-archival and therefore not a prior publication of these results. The differences are at the component level (Section 5).

Figure 2. In an earlier visualisation we compared against an early topological variant; when that panel was dropped, the name of our own method was deleted by mistake and the older label was left behind. The figure content has always been ours. The correct label is TopoReg, and the figure has also been redrawn with the point sets distinguished by colour and a labelled colour scale.

Citations. The duplicated entries (CPD, Carrière et al., Chen et al.) have been merged; thin-plate spline interpolation is now cited; and the manifold-learning lineage, including density-based metric learning and spectral-method formulations of persistent homology, has been added to the related work.

2. The control experiment and the attribution

On Lung, the comparison that isolates the descriptor is against a PointPWC model that keeps the same edge-length loss but omits both the descriptor and FiLM. Its parameter count equals the PointPWC baseline exactly (7,720,236), so no topological signal enters the network, while the loss stays active at weight 0.05.

Protocol: 1,500 epochs, batch size 4, identical augmentation and evaluation pipeline; entries are landmark TRE in mm.

ConfigurationDescriptorLtopo\mathcal{L}_{\text{topo}}Run 1Run 2Run 3Mean ± std
PointPWC + Ltopo\mathcal{L}_{\text{topo}} (control)noyes2.312.222.282.270 ± 0.037
TopoReg (full)yesyes2.182.202.182.187 ± 0.009
Difference0.083

Because the auxiliary loss is present in both configurations, the comparison isolates the conditioning branch. The gap is about 2.2× the control's spread and about 9× that of the full model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.