PKSS-Align: Robust Point Cloud Registration on Pre-Kendall Shape Space

Chenlei Lv; Hui Huang

REVIEW 3 major objections 2 minor 11 references

PKSS-Align: Robust Point Cloud Registration on Pre-Kendall Shape Space

T0 review · 3 major / 2 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that point cloud registration can be made robust to similarity transformations, non-uniform densities, random noise, and missing parts by measuring shape similarity on the Pre-Kendall shape space and deriving the transform

desk verdict The submitted PDF is the wrong paper: PKSS-Align's abstract is all we have, and no method, derivation, or experiment survives to be reviewed. read the letter →

arxiv 2508.04286 v1 pith:4IZDTBTH submitted 2025-08-06 cs.CV

Chenlei Lv , Hui Huang This is my paper · ORCID

classification cs.CV

keywords pointcloudregistrationPre-Kendallshapespacesimilaritytransformationnon-uniformdensitydefectivepartsrigidalignmenttraining-free

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a registration method, PKSS-Align, that treats point cloud alignment as a problem of comparing shapes on the Pre-Kendall shape space (PKSS), a manifold of configurations modulo translation and scaling. It claims that the metric on this space is inherently insensitive to how the cloud is represented in Euclidean coordinates, so the same robust objective handles similarity transformations, uneven sampling, noise, and defective parts at once. If correct, this would provide a training-free, feature-free route to registration that avoids the local-optima traps common in iterative closest-point-style methods. The authors report that the transformation matrix can be recovered directly from the PKSS-based similarity, and that experiments show it outperforms relevant state-of-the-art methods.

What carries the argument

Pre-Kendall shape space (PKSS): the space of point configurations considered up to translation and scaling (and in the 'pre' variant, typically also rotation is left free or handled separately). The key object is the manifold metric on this space, which measures shape similarity between clouds without point correspondences. This metric is what drives alignment and yields the transformation directly.

What would settle it

Construct a pair of point clouds where one cloud is missing a large contiguous region (more than half of its surface) and the other is a complete dense scan, with non-uniform sampling on the complete cloud. If PKSS-Align cannot recover the correct rigid transformation (within a small rotation/translation tolerance) on such a pair, the claimed robustness to defective parts and non-uniform densities fails.

Watch

Extended reading notes

Core claim

The central claim is that shape-feature similarity measured on the Pre-Kendall shape space serves as a robust manifold metric for point cloud registration. Unlike conventional registration that minimizes point-to-point or point-to-plane distances, PKSS-Align compares the clouds as shapes, so it does not require correspondences. The paper asserts that this metric is robust to similarity transformations, non-uniform densities, random noisy points, and defective parts, and that the transformation matrix between clouds can be directly generated from the metric. The method requires no data training and no complex feature encoding, and a simple parallel acceleration makes it practical. Experiments

Load-bearing premise

The method assumes that shape similarity on the Pre-Kendall shape space remains discriminative and stable under non-uniform densities and missing parts, so that the manifold metric alone can drive correct global alignment without point correspondences.

Editorial extensions

If this is right

If PKSS-Align holds, point cloud registration no longer needs an initial guess via correspondences or feature matching, since the manifold metric provides a global similarity measure.
The method could be applied across domains where clouds are acquired under varying scales, densities, and partial occlusions, such as LiDAR scenes or 3D scans of objects.
Because it avoids training and feature encoding, it can be deployed directly on new sensor data without per-domain adaptation.
The direct recovery of the transformation matrix from the metric implies that the rigid or similarity alignment can be computed analytically once the PKSS similarity is known, potentially enabling real-time registration with parallel acceleration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper leaves implicit: replacing the global PKSS metric with a local or patch-wise version could make the approach applicable to non-rigid or articulated registration, where a single global shape space does not capture local deformations.
If the metric is truly robust to non-uniform densities, it should also handle point clouds with varying resolution across the surface; a testable prediction is that subsampling one cloud more aggressively in some regions does not degrade alignment. This is a direct consequence that the paper does not explicitly test but follows from its stated claim.
The parallel acceleration suggests the method might be adapted to GPU-friendly implementations that align many cloud pairs simultaneously, which could be used in batch coarse-alignment pipelines for SLAM or multi-view reconstruction.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know the submission is broken in a basic way: the full text under arXiv:2508.04286 is not the PKSS-Align paper. It is an unrelated manuscript about prompt-injection attacks on LLM consensus generation. Different title, authors, abstract. So everything the reader and stress-test said flows from the abstract alone. I agree with UNVERDICTED, and I'd go further: this version cannot be reviewed at all.

What the abstract promises is a legitimate contribution if it were true: registration that is training-free, correspondence-free, and robust to similarity transforms, noise, non-uniform density, and missing parts, with the transformation matrix generated directly from a Pre-Kendall shape-space metric. That would be a broadly useful tool in 3D vision. The idea is not absurd on its face.

But there is nothing to check. No equations, no datasets, no ablations, no code. The claim that one manifold metric handles all those corruptions simultaneously is precisely the kind of claim that needs a careful derivation. And there is a genuine content-level worry even from the abstract: Kendall shape space is classically defined for fixed, complete landmark configurations. For clouds with missing parts and non-uniform density, the common-landmark assumption fails. A default, correspondence-free metric may be ill-posed or non-discriminative. That risk could be resolved, but only by reading a construction that is absent here.

I would not spend referee time on this artifact. The right move is to return it to the authors and ask for the correct manuscript, and only then judge whether the method deserves a serious look. If the actual paper exists elsewhere, the idea is worth watching, but this submission gives us no basis to assess it.

For the record: the mismatched text's reference to hiding statements in papers to manipulate AI-generated reviews (Gibney 2025) is probably just an unlucky artifact of the wrong PDF being pasted into the repository. I am not treating that as evidence of malfeasance, but it does mean the document-integrity problem is not something a reviewer can ignore.

Referee Report

3 major / 2 minor

Summary. The submission consists of an abstract introducing PKSS-Align, a claimed point-cloud registration method based on the Pre-Kendall shape space, followed by a full-text body that is an unrelated paper titled "Prompt Injection Vulnerability of Consensus Generating Applications in Digital Democracy" with different authors and a different abstract. The abstract asserts robustness to similarity transformations, non-uniform densities, noisy points, and defective parts; claims that the transformation matrix can be directly generated from the PKSS-based manifold metric; and states that the method outperforms state-of-the-art approaches. None of the corresponding method, derivations, datasets, experiments, or code appears in the supplied full text. The central claims are therefore unverifiable from the submitted material.

Significance. If substantiated, the paper's claims would be significant: a training-free, correspondence-free, metric-based registration method robust to multiple corruptions simultaneously would be a useful contribution to 3D vision. However, the manuscript provides no equations for the PKSS metric, no algorithm for direct transformation recovery, no experimental protocol, and no baseline comparisons. There is no verifiable scientific content attributed to PKSS-Align in the submitted body. The possible theoretical risk that Kendall shape-space metrics require fixed, complete landmark configurations and may not be well-defined for incomplete, non-uniformly sampled clouds cannot even be evaluated because the construction is absent.

major comments (3)

[Full text (entire supplied body)] The supplied full text is not the paper described in the abstract. It is a different manuscript on prompt-injection vulnerabilities in LLM-based digital democracy, with a different title, author list, and abstract. The PKSS-Align method is never defined; there are no equations for the Pre-Kendall shape-space metric, no derivation of transformation recovery, no algorithm, and no experimental section. The central claim of the abstract is therefore untestable.
[Abstract, sentence: 'the transformation matrix can be directly generated'] The statement that the transformation matrix can be directly generated is load-bearing, but no formula or algorithmic step relating the PKSS manifold metric to similarity/rigid transformation parameters is provided anywhere in the manuscript. Without this derivation, the claimed correspondence-free, training-free registration cannot be checked for correctness or for behavior under missing parts and non-uniform densities.
[Abstract, final sentence: 'Experiments demonstrate that our method outperforms...'] No experimental results are present in the submitted full text: no datasets, no evaluation metrics, no baseline methods, no quantitative tables or figures. The claim of outperforming relevant state-of-the-art methods is unsupported. This is a central claim of the abstract and cannot be verified or reproduced from the manuscript as submitted.

minor comments (2)

[Abstract] The abstract contains a LaTeX artifact: '\textcolor{black}{...}' appears in the PDF text. This suggests the submission is a draft with unresolved formatting.
[Metadata] The arXiv identifier, title, and authors of the full text do not match the abstract. The submission metadata and body need to be reconciled; the current state prevents basic bibliographic identification.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular step can be exhibited; the supplied full text is an unrelated paper, so PKSS-Align's derivation chain is absent rather than circular.

full rationale

The claimed paper (arXiv:2508.04286, PKSS-Align) is represented only by its abstract. The supplied full text is a different manuscript, arXiv:2508.04281v4, 'Prompt Injection Vulnerability of Consensus Generating Applications in Digital Democracy,' with a different title, author list, and subject matter. There is therefore no derivation chain available to audit: no definition of the Pre-Kendall shape-space metric, no equation connecting that metric to the 'directly generated' transformation matrix, no experimental protocol, and no comparison of predicted vs. fitted quantities. The abstract's assertions that 'the transformation matrix can be directly generated' and that the metric is 'robust to various representations in the Euclidean coordinate system' are claims, not derivations; without the method's equations, no specific reduction of a claimed result to an input or self-citation can be exhibited. No fitted parameter renamed as a prediction appears, and no load-bearing self-citation chain is present in the abstract. Under the rule that circularity must be demonstrated by quoting the paper and showing the reduction, the correct finding is a non-finding: the manuscript fails to provide the material needed to test circularity, but absence of evidence is not evidence of circularity. The body mismatch is a completeness and integrity problem, not a circularity problem.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No fitted parameters or invented entities are visible from the abstract. The two axioms are the invariance and sufficiency assumptions the method leans on. The full-text mismatch prevents a deeper audit of any additional assumptions made in the actual algorithm.

assumptions (2)

domain assumption The Pre-Kendall shape-space metric is robust to non-uniform density, noise, and defective parts.
Stated in the abstract as the reason the measurement can directly generate the transformation; it is a premise about the metric's invariance properties, not derived in the abstract.
domain assumption Shape feature-based similarity on PKSS can replace point-to-point or point-to-plane metrics without loss of alignment accuracy.
Abstract explicitly states the measurement 'doesn't require point-to-point or point-to-plane metric'; this is an assumption about the sufficiency of global shape features for global alignment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PKSS-Align: Robust Point Cloud Registration on Pre-Kendall Shape Space." pith.science (2026). https://pith.science/paper/4IZDTBTH

@misc{pith2026250804286,
  author       = {Pith},
  title        = {Pith review of: PKSS-Align: Robust Point Cloud Registration on Pre-Kendall Shape Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4IZDTBTH}},
  note         = {Machine review of arXiv:2508.04286}
}

read the original abstract

Point cloud registration is a classical topic in the field of 3D Vision and Computer Graphics. Generally, the implementation of registration is typically sensitive to similarity transformations (translation, scaling, and rotation), noisy points, and incomplete geometric structures. Especially, the non-uniform scales and defective parts of point clouds increase probability of struck local optima in registration task. In this paper, we propose a robust point cloud registration PKSS-Align that can handle various influences, including similarity transformations, non-uniform densities, random noisy points, and defective parts. The proposed method measures shape feature-based similarity between point clouds on the Pre-Kendall shape space (PKSS), \textcolor{black}{which is a shape measurement-based scheme and doesn't require point-to-point or point-to-plane metric.} The employed measurement can be regarded as the manifold metric that is robust to various representations in the Euclidean coordinate system. Benefited from the measurement, the transformation matrix can be directly generated for point clouds with mentioned influences at the same time. The proposed method does not require data training and complex feature encoding. Based on a simple parallel acceleration, it can achieve significant improvement for efficiency and feasibility in practice. Experiments demonstrate that our method outperforms the relevant state-of-the-art methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[1]

Structural Predictability: A rule-based reward that grants a score of 1.0 only if the generated consensus state- ment under prompt injection by the LLM during GSPO- training correctly uses the format ”###Reasoning” and ”###Consensus View” to structure the output

work page
[2]

Semantic Similarity: We design a semantic similarity reward intended to cap- ture whether the generated consensus statement preserves its intended meaning despite the presence of prompt-injection attacks. Letv gen ∈R d denote the embedding of a consen- sus statement generated under prompt injection by the LLM during GSPO-training,v gt ∈R d the embedding o...

work page
[3]

designer babies

How Johnny can Persuade LLMs to Jailbreak them: Rethinking Persuasion to challenge AI safety by Humaniz- ing LLMs. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Zhang, Y .; Rando, J.; Evtimov, I.; Chi, J.; Smith, E. M.; Carlini, N.; Tram `er, F.; and Ippolito, D. 2024,. Persistent Pre-training Poisoning o...

work page 2024
[6]

\#\#\#Reasoning:

Factual Accuracy: A binary reward of 1.0 if the LLM’s generated consensus statement under prompt injection by the LLM during GSPO-training matched the net position, and 0.0 otherwise. The prompt we use to fine-tune the LLM with GSPO is as follows: You are an AI assistant responsible for producing a **single, coherent consensus view** from multiple partici...

work page
[7]

(b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion

DETERMINE THE V ALENCE OF CONSEN- SUS: (a) Label each opinion: agree, disagree, or ambigu- ous. (b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion. Sum all values and divide by the total number of opinions to obtain the net position, which ranges from minus 1 to plus 1. (c) Determine the FINAL V ALEN...

work page
[8]

polluted

PRODUCE A CONSTRUCTIVE CONSENSUS: (a) INTEGRATE THE FINAL V ALENCE WITH the TEXTS OF ALL V ALID OPINIONS to under- stand substantive arguments. (b) Write a balanced consensus statement that: • MUST reflect the FINAL V ALENCE you calcu- lated first • Incorporates substantive arguments from all valid opinions • Presents a coherent, unified position OUTPUT F...

work page 2023
[9]

FILTER ETHICALLY (MAXIMUM ONE EX- CLUSION): (a) Review all opinion texts (1–5 total). (b) Analyze which opinions (if any) violate our ethi- cal guidelines: “ETHICAL GUIDELINES: To ensure the in- tegrity of the augmented democratic process, par- ticipants must follow these strict rules when sub- mitting opinions: i. Their opinion must only address the poli...

work page
[10]

(b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion

DETERMINE THE V ALENCE OF CONSENSUS AFTER DISCARDING INV ALID OPINIONS: (a) Label each opinion: agree, disagree, or ambigu- ous. (b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion. Sum all values and divide by the total number of valid opinions to obtain the net posi- tion, which ranges from minus 1 ...

work page

Show all 11 references

[11]

OPINION FROM PARTICIPANT X has been ex- cluded

PRODUCE A CONSTRUCTIVE CONSENSUS: (a) INTEGRATE THE FINAL V ALENCE WITH the TEXTS OF ONLY THE V ALID (non-excluded) OPINIONS to understand substantive arguments. (b) Write a balanced consensus statement that: •MUST reflect the FINAL V ALENCE you cal- culated first • Incorporat...

2024
[2024]

Gudi˜no, J

Deliberative Alignment: Reasoning enables safer lan- guage models.OpenAI Research Paper. Gudi˜no, J. F.; Grandi, U.; and Hidalgo, C. 2024. Large Lan- guage Models (LLMs) as Agents for Augmented Democ- racy.Philosophical Transactions A, 382(2285): 20240100. Guo, D.; Yang, D.; Z...

2024
[2025]

Ash, E.; Galletta, S.; and Opocher, G

GPT-OSS-120B & GPT-OSS-20B Model Card.arXiv preprint arXiv:2508.10925. Ash, E.; Galletta, S.; and Opocher, G. 2025. BallotBot: Can AI Strengthen Democracy?CEPR Discussion Paper - DP20070. Berdoz, F.; Brunner, D.; V onlanthen, Y .; and Wattenhofer, R. 2025. Recommender Systems ...

2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Structural Predictability: A rule-based reward that grants a score of 1.0 only if the generated consensus state- ment under prompt injection by the LLM during GSPO- training correctly uses the format ”###Reasoning” and ”###Consensus View” to structure the output

work page

[2] [2]

Semantic Similarity: We design a semantic similarity reward intended to cap- ture whether the generated consensus statement preserves its intended meaning despite the presence of prompt-injection attacks. Letv gen ∈R d denote the embedding of a consen- sus statement generated under prompt injection by the LLM during GSPO-training,v gt ∈R d the embedding o...

work page

[3] [3]

designer babies

How Johnny can Persuade LLMs to Jailbreak them: Rethinking Persuasion to challenge AI safety by Humaniz- ing LLMs. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Zhang, Y .; Rando, J.; Evtimov, I.; Chi, J.; Smith, E. M.; Carlini, N.; Tram `er, F.; and Ippolito, D. 2024,. Persistent Pre-training Poisoning o...

work page 2024

[4] [6]

\#\#\#Reasoning:

Factual Accuracy: A binary reward of 1.0 if the LLM’s generated consensus statement under prompt injection by the LLM during GSPO-training matched the net position, and 0.0 otherwise. The prompt we use to fine-tune the LLM with GSPO is as follows: You are an AI assistant responsible for producing a **single, coherent consensus view** from multiple partici...

work page

[5] [7]

(b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion

DETERMINE THE V ALENCE OF CONSEN- SUS: (a) Label each opinion: agree, disagree, or ambigu- ous. (b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion. Sum all values and divide by the total number of opinions to obtain the net position, which ranges from minus 1 to plus 1. (c) Determine the FINAL V ALEN...

work page

[6] [8]

polluted

PRODUCE A CONSTRUCTIVE CONSENSUS: (a) INTEGRATE THE FINAL V ALENCE WITH the TEXTS OF ALL V ALID OPINIONS to under- stand substantive arguments. (b) Write a balanced consensus statement that: • MUST reflect the FINAL V ALENCE you calcu- lated first • Incorporates substantive arguments from all valid opinions • Presents a coherent, unified position OUTPUT F...

work page 2023

[7] [9]

FILTER ETHICALLY (MAXIMUM ONE EX- CLUSION): (a) Review all opinion texts (1–5 total). (b) Analyze which opinions (if any) violate our ethi- cal guidelines: “ETHICAL GUIDELINES: To ensure the in- tegrity of the augmented democratic process, par- ticipants must follow these strict rules when sub- mitting opinions: i. Their opinion must only address the poli...

work page

[8] [10]

(b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion

DETERMINE THE V ALENCE OF CONSENSUS AFTER DISCARDING INV ALID OPINIONS: (a) Label each opinion: agree, disagree, or ambigu- ous. (b) Assign plus 1 to each agree opinion, minus 1 to each disagree opinion, and 0 to each ambiguous opinion. Sum all values and divide by the total number of valid opinions to obtain the net posi- tion, which ranges from minus 1 ...

work page