{"id":"1438a96d-c559-47e2-bff6-6e1d04830e93","arxiv_id":"2505.04485","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Frame Averaging applied to KPConv yields FA-KPConv, a parameter-free wrapper that makes KPConv networks exactly Euclidean invariant or equivariant, improving robustness on rotated and low-data benchmarks.","lead":"The authors wrap the popular KPConv point-cloud network in a frame-averaging wrapper that makes it exactly invariant or equivariant to rotations, translations, and reflections. The resulting FA-KPConv keeps the same number of parameters as the baseline and shows better accuracy on rotated test data and with little training data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15) is correct under the joint action, but the 'simply wrapping' and 'no input information compromise' claims fail for scalar or non-geometric features; the experiments alter the baseline to make features geometric, so the practical claim is not established.","rationale":"I read the paper in good faith and agree that the method can deliver exact invariance for networks whose input features are geometric vectors transforming under the same group element. The mathematical construction in Eqs. (14)-(15) is sound; the reader's concern about an unproven joint-action property is actually resolved by the standard frame equivariance condition, so I do not share that specific worry. However, the paper's central claim is broader: it advertises a simple wrapper that preserves input information. That claim requires the input features to already satisfy the joint-action assumption, which is not true for the common KPConv setup with scalar features. The classification experiments explicitly modify the feature representation from a scalar constant to a 3-vector and then treat it as rotating, so the comparison is against an altered baseline. The registration experiments wrap only the KPConv backbone, not the full GeoTransformer, so the exact invariance guarantee does not apply to the whole model. These gaps mean the practical scope is narrower than stated, and the paper should either restrict its claims or provide experiments with natural geometric features and the original baselines. The parameter-count preservation and the plausible gains on rotated/low-data benchmarks are real strengths, and the method is useful in the restricted geometric-feature setting, so the verdict should remain conditional rather than rejection.","tokens_in":11641,"tokens_out":12770,"duration_ms":137400,"concrete_test":"Compute Eq. (15) for a single KPConv layer with a scalar feature replicated to 3D, using a fixed point cloud X and a rotated version gX with the feature unchanged. Compare the two outputs; if they differ, exact invariance under the natural coordinate-only rotation fails. If they agree, empirically evaluate a KP-CNN with original scalar constant features wrapped by FA-KPConv on rotated ModelNet40 with unchanged features to test the 'no compromise' claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The reader's stated concern about Eq. (15) does not land: for any equivariant frame F (i.e., F(gX)=gF(X)), averaging over F(X) with the joint action (X,Fin)->(gX,gFin) yields exact invariance. This follows directly from the frame averaging theorem and requires no 'extensive testing'; the paper's missing proof is a presentation gap, not a correctness gap. The load-bearing problem is different: the invariance is for the joint action, which is not the natural action for many point-cloud features. For scalar features (constant 1, intensity, color), a rotation acts trivially on Fin; the paper instead requires cin=kd and reshapes/replicates features into d-vectors so that Fin transforms by g. This changes the input representation and is not a mere wrap. In the classification experiments, the original KP-CNN's constant scalar feature is replaced by a constant 3-vector that the group rotates, and the baseline is the modified network, not the original KP-CNN. Thus the central claim that FA-KPConv 'preserves the number of learnable parameters and does not compromise any input information' is not supported for non-geometric features. The exact invariance holds only when features are geometric vectors sharing the group action, a condition not stated in the abstract and not satisfied by many practical KPConv uses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FA-KPConv, a wrapper around KPConv-based point-cloud networks that uses Frame Averaging (Puny et al.) to make the network exactly invariant or equivariant to translations, rotations, and reflections. The construction represents a network as Y = f(X, Fin), computes a frame from the covariance/eigendecomposition of the coordinates X, and replaces f by an average over the frame, after reshaping input features into d-dimensional blocks so that they transform under the same Euclidean action. Experiments on ModelNet40 classification and 3DMatch/3DLoMatch registration compare modified KPConv baselines using constant 3-vector features against their FA versions, reporting gains mainly on rotated test data and in the low-training-data regime.","tokens_in":11925,"tokens_out":7746,"duration_ms":87144,"significance":"If properly scoped, the contribution is useful: it provides a simple recipe for making KPConv-based networks exactly invariant under Euclidean motions for geometric vector features, with no additional trainable parameters, and the experiments support gains in the low-data and rotated-test regimes. The mathematical claim is correct under the joint action on X and Fin, contrary to one of the stress-test concerns, but the paper states this without proof and overstates the 'simple wrapping' and 'no compromise' aspects for non-geometric features. The rotated-test protocol is also ambiguous about whether features are transformed. These issues are fixable and do not invalidate the core construction, but they do affect the exactness claim as advertised.","major_comments":[{"comment":"The paper asserts that exact invariance holds when f depends on both X and Fin, but justifies this only with 'We verified this via extensive testing' in Section III-C. This is load-bearing because exactness is the central claim. Provide a short proof: for an equivariant frame satisfying F(hX) = hF(X), re-indexing g = h g' in the sum over F(hX) gives the claimed equality in Eq. (15); the equivariant case Eq. (14) follows similarly. The proof should also state explicitly that Fin is acted on by the same g; without that joint action, Eqs. (16)-(17) do not follow.","section":"III-C (Eqs. 14-17)"},{"comment":"The abstract's claim of 'simply wrapping around an existing KPConv-based network' is not accurate for the networks used in the experiments. The original KP-CNN and GeoTransformer use a constant scalar feature equal to 1 (cin = 1), which is invariant under rotations; the paper changes the input to a constant 3-vector and compares FA models against this modified baseline, not against the original network. Thus the experiments do not show that the original KPConv network can be wrapped without changing its input representation, and the claim 'not compromising any input information' is only true in the weak sense of preserving information content, not in the sense of preserving the original feature space. Please state the scope explicitly: exact invariance applies when Fin consists of geometric d-vectors that transform under the same Euclidean action.","section":"IV-A, IV-B and Abstract"},{"comment":"The rotated-test protocol is ambiguous: Section IV says the test data are 'once on a rotated version of it, in which each sample is randomly rotated' but does not state whether the input features Fin (the vector-ones features) are rotated together with the coordinates X. Since Eq. (17) guarantees invariance only for the joint action (gX, gFin), the exact-invariance claim is not meaningfully evaluated unless the same g is applied to Fin in the rotated tests. Specify the exact transformation applied to X and Fin in the rotated experiments, and for registration state whether the source and target point clouds receive the same or independent random rotations.","section":"IV (Experimental protocol)"}],"minor_comments":[{"comment":"Please clarify the reshaping used to apply g to Fin: the sentence requiring cin to be a multiple of d should explicitly define the reshape/respace operation and its inverse, since Eq. (14) uses the notation g·Fin without a formal definition.","section":"III-C"},{"comment":"Consider adding the results of the original unmodified KP-CNN with scalar constant features (cin = 1) to Table II, so readers can quantify the effect of the vector-one reparameterization separately from the effect of frame averaging.","section":"Table II"},{"comment":"The conjecture that FA models 'spend capacity on learning to ignore the invariances imposed by design' is speculative; it would be helpful to cite or propose a concrete diagnostic, such as probing the frame-averaged outputs on aligned versus rotated inputs.","section":"V (Analysis)"},{"comment":"The two diagrams in Figure 1 are nearly identical and the labels are small; enlarging the figure and highlighting the difference between the equivariant and invariant cases would improve readability.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The core construction is mathematically sound under the joint action on X and Fin, and the experiments are a useful proof of concept for geometric features. The main concerns for publication are the missing proof for Eqs. (14)-(15), the overstatement of 'simple wrapping' for networks with scalar input features, and the ambiguous rotated-test protocol. These are addressable in revision, but the current abstract and claims overstate the scope. I do not see a novelty-disclosure problem: frame averaging is cited from prior work, and the contribution is the KPConv integration and its evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core thing to know: this is a legitimate integration of Puny et al.'s frame averaging into KPConv, and the central invariance claim is mathematically correct. The reader's worry about Eq. (15) doesn't really land—for an equivariant frame F, averaging over F(X) under the joint action (gX, gFin) gives exact invariance. That follows directly from the standard FA theorem; the missing proof is a presentation gap, not a correctness gap. The paper's more serious weakness is how it handles input features. The method requires cin to be a multiple of d and reshapes/replicates features so they rotate with the cloud. That is fine for coordinate-like features, but wrong for scalar or color features, where rotating the feature values is semantically meaningless. The experiments sidestep this by replacing the constant 1 of KP-CNN with a constant 3-vector and comparing against a baseline that also uses that vector. So the claim that FA-KPConv is a 'simple wrap' with 'no input information compromise' only holds for geometric features; for non-geometric features you would need to leave Fin untransformed (which would still give invariance), but the paper doesn't say that. On the empirical side, the gains on rotated data and in the low-data regime are real and consistent, and the paper is honest about the tradeoff on canonically aligned data. But there are no error bars, no multiple seeds, and no code yet, so the exact magnitude of the improvements is uncertain. Overall this is a useful practical result for a widely used backbone, with correct math and an oversold presentation. A serious referee can fix this by asking for a proof or precise statement of the joint-action invariance, a discussion of non-geometric features, and ideally error bars and a release of the code. I'd send it to review, and I'd expect a conditional acceptance after revision.","headline":"Sound FA-on-KPConv engineering, but the 'simply wrapping' slogan oversells it; the joint-action invariance is correct, the feature-handling story is not.","tokens_in":12430,"tokens_out":5204,"would_cite":true,"duration_ms":52294,"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":"Wrapping KPConv-based point-cloud networks in frame averaging yields exact invariance and equivariance to Euclidean transformations without adding any learnable parameters.","keywords":["Frame Averaging","KPConv","point cloud classification","point cloud registration","equivariance","invariance","Euclidean group","3D deep learning"],"falsifier":"Take a KPConv network with non-geometric input features, apply a random rotation to both the coordinates and the features, and check whether the frame-averaged output is exactly unchanged; any rotation that changes the output would falsify the claimed exact invariance for the general f(X, Fin) case. A second test is to use point clouds with degenerate covariance spectra, such as planar or collinear points, where the eigendecomposition frame is not unique and the averaging may become fragile.","tokens_in":11454,"feed_emoji":"🔄","tokens_out":6593,"duration_ms":67441,"temperature":0.7,"pith_summary":"This paper claims that any point-cloud network built from KPConv layers can be made exactly invariant or equivariant to translations, rotations, and reflections by wrapping it in Frame Averaging, without adding a single learnable parameter. The construction replaces the intractable average over the whole Euclidean group with an average over a small input-dependent frame: the orthogonal frame obtained from the eigendecomposition of the point cloud's covariance matrix. The authors argue that this embeds geometric prior knowledge directly into the network, so classification and registration improve precisely in the regimes where such priors matter, namely rotated test data and scarce training data, while performance on already-aligned data stays roughly unchanged. The paper demonstrates this on ModelNet40 classification and 3DMatch/3DLoMatch registration by comparing KPConv baselines with their FA-KPConv counterparts.","feed_headline":"Frame averaging makes KPConv rotation-proof with zero extra parameters","feed_subtitle":"The same number of trainable parameters still handles randomly rotated scans and low-data training sets.","key_machinery":"The load-bearing object is the frame F(X): for a point cloud X, one computes the centroid c and covariance matrix C, then takes the eigendecomposition C = Q\\Lambda Q^T, whose unit eigenvectors define an orthogonal frame up to sign choices. The set of all sign choices, together with the centering translation, forms a frame for the desired group, with size 1 for translations, 4 for rotations, 8 for rotations and reflections, and similarly for their combinations with translations. Frame averaging then replaces the intractable group average with an average over this small set, keeping the symmetrization exact. The paper wraps KPConv networks with this averaging, requiring the feature dimension to be a multiple of 3 so that $g^{{-1}}$ can act on Fin, and using a constant vector 1 in $R^{3}$ as the baseline input feature.","core_discovery":"The central claim is that Frame Averaging can be applied directly to a KPConv network f(X, Fin), where X are point coordinates and Fin are input features, to produce a function that is exactly invariant or equivariant under simultaneous Euclidean transformations of X and Fin. Specifically, the paper defines the equivariant function \\hat f and the invariant function \\bar f by averaging f over the frame F(X) of the input cloud, as in equations (14) and (15). It claims that this symmetrization remains exact even when the network depends on both coordinates and features, not only on coordinates, and that this holds for any KPConv-based architecture, from a single convolution layer to full classification or registration networks. The cost is not in parameters, which stay unchanged, but in compute and memory, which grow by a factor of |F(X)|, up to 8 for the full Euclidean group in 3D.","pith_inferences":["A formal proof that the frame-averaged function remains exact for f(X, Fin) would require verifying that the frame is equivariant, F(gX) = gF(X), and that the feature action is linear; the paper's 'extensive testing' leaves this unproven but testable.","If exactness holds for arbitrary features, the same wrapping could be applied to other coordinate-based 3D backbones beyond KPConv, yielding exactly equivariant versions of sparse convolutions or graph networks at the same parameter count.","The paper's baseline replaces scalar input features with a constant 3-vector; a natural extension would test what happens when Fin contains non-geometric information such as color or semantic labels, where the symmetry action on features is not physically meaningful.","The |F(X)| compute multiplier could be reduced by selecting one canonical eigenvector orientation instead of averaging over all sign choices, at the price of exactness when eigenvalues are degenerate."],"forward_implications":["FA-KPConv models trained on a small fraction of ModelNet40 remain almost unaffected by random test-time rotation: with 9843 training samples, accuracy on rotated data is 87.0% versus 87.1% on the original data, while the baseline drops from 90.4% to 44.6%.","On the low-overlap 3DLoMatch benchmark with 1k training samples, FA-GeoTransformer improves inlier ratio by up to 19.7% over the baseline and reduces relative rotation error by about 9.9% on rotated test data.","Because no parameters are added, any performance gain must come from the embedded geometric prior rather than from increased model capacity.","In already-aligned, canonically oriented datasets the enforced invariance can slightly hurt, which the paper attributes to the model spending capacity to work around an unnecessary constraint."],"supporting_citations":[{"why":"Defines the KPConv convolution and the KP-CNN/KP-FCNN architectures that the paper wraps.","marker":"[1]"},{"why":"Introduces frame averaging, the frames for Euclidean motion groups, and the exactness argument for functions of coordinates.","marker":"[8]"},{"why":"Provides GeoTransformer, the KPConv-based registration architecture used as the registration baseline.","marker":"[6]"},{"why":"Supplies the ModelNet40 benchmark used for the classification experiments.","marker":"[42]"},{"why":"Supplies the 3DMatch and 3DLoMatch benchmarks and metrics used for the registration experiments.","marker":"[43]"}],"fun_headline_variants":["Exact rotation invariance for KPConv via frame averaging, no new parameters","Zero extra parameters: FA-KPConv makes KPConv exactly equivariant to rotations","FA-KPConv: frame averaging gives KPConv exact Euclidean symmetry, no extra weights","No new parameters: FA-KPConv makes KPConv exactly invariant to rotations, reflections"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that frame averaging remains exactly invariant or equivariant when the network consumes both coordinates X and input features Fin, a property the paper asserts was 'verified via extensive testing' without providing a proof or test details; if this premise fails, the advertised exact symmetry does not hold for networks that use features beyond coordinates.","fun_headline_variants_meta":{"raw":{"variants":["Exact rotation invariance for KPConv via frame averaging, no new parameters","Zero extra parameters: FA-KPConv makes KPConv exactly equivariant to rotations","FA-KPConv: frame averaging gives KPConv exact Euclidean symmetry, no extra weights","No new parameters: FA-KPConv makes KPConv exactly invariant to rotations, reflections"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000504,"raw_usage":{"total_tokens":2434,"prompt_tokens":891,"completion_tokens":1543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":1456}},"tokens_in":507,"tokens_out":1543,"duration_ms":11308,"temperature":1.0,"reasoning_tokens":1456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:27:17.434083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a KPConv network with non-geometric input features, apply a random rotation to both the coordinates and the features, and check whether the frame-averaged output is exactly unchanged; any rotation that changes the output would falsify the claimed exact invariance for the general f(X, Fin) case. A second test is to use point clouds with degenerate covariance spectra, such as planar or collinear points, where the eigendecomposition frame is not unique and the averaging may become fragile.","supporting_citations":[{"cited_title":"Kpconv: Flexible and deformable convolution for point clouds,","cited_arxiv_id":null,"evidence_quote":"Defines the KPConv convolution and the KP-CNN/KP-FCNN architectures that the paper wraps."}],"review_version":1}