{"id":"3bb7a7ed-c053-41ae-9bba-cbf3cf38435d","arxiv_id":"2505.04540","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A Gaussian-weighted ICP variant is claimed to align point clouds under full-range rotations and non-Gaussian noise, yet the convergence proof is unproven and several reported error statistics violate basic inequalities.","lead":"The paper proposes a modified ICP algorithm, called ESM-ICP, that weights point correspondences with a Gaussian exponential similarity matrix to align 3D point clouds under large rotations and noise. The authors report it outperforms classical and learning-based methods in synthetic tests on ModelNet40 and Stanford Bunny, but the support contains numerical and logical inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 100%-success robustness claim rests on an invalid convergence theorem and a same-point-set protocol that avoids partial overlap; the abstract's practical claim is therefore unsupported.","rationale":"The reader correctly identified the synthetic exact-correspondence setup as the weakest assumption. My pass confirms this and adds that the only theoretical support, Theorem 1, cannot rescue the claim because its monotonicity step is invalid: the algorithm re-establishes correspondences and weights at every iteration, so E_{k+1} is not the objective minimized at step k. This is a correctness risk, not a disagreement with field consensus: local reweighted-ICP methods are not generally global, and the paper provides no basin-of-attraction analysis. The empirical 100% success rate, if reproduced, would be remarkable, but as reported it is not checkable because no per-trial success threshold or distribution is given, and the exact-correspondence protocol removes the partial-overlap failure mode central to real multi-view registration. The KISS-Matcher comparison (Section V-D) actually shows an outlier case where ESM-ICP has slightly higher rotation MSE than KISS-Matcher (0.073224 vs 0.073004 in Table VI), further undercutting the 'consistently outperformed' wording, though that is secondary. I would keep the reader's REJECT verdict. A revision could address this by adding a correct convergence analysis and evaluating on partial-overlap benchmarks such as subsampled or multi-view scans, with per-trial results.","tokens_in":14801,"tokens_out":11508,"duration_ms":118978,"concrete_test":"Run the public ESM-ICP implementation on ModelNet40 using the paper's protocol: source = a random rotation of the target in [-pi, pi], sigma per stated density, at most 100 iterations, with a pre-specified success threshold such as RMSE(R) < 1e-4 and RMSE(t) < 1e-3, for 100 fresh trials. Then repeat with the source formed by subsampling the target to 70% of its points so exact correspondences do not exist. If the success rate drops below 100% in either condition, the 'aligned all test pairs' claim and its transfer to real scans are falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-B reports that over 2,000 random rotations in [-pi, pi] ESM-ICP aligned all test pairs within 100 iterations, and the abstract generalizes this to practical registration of multi-view scans. The theoretical support for this claim is Theorem 1 (Section IV-A), but the proof is invalid: after iteration k the correspondences c(i) and weights w_i are recomputed from the new transform, so E_{k+1} in Eq. (12) is a different objective than E_k; A_{k+1} minimizes E_k, not E_{k+1}, and the asserted inequality E_{k+1} < E_k in Eq. (15) does not follow. Even a valid monotone scalar error sequence would not imply convergence to the global optimum for this nonconvex nearest-neighbor iteration. The only experiments use the Section V protocol in which Source is a rigidly transformed copy of Target, so every source point has an exact target match and cardinalities are equal; partial overlap, different sampling density, and viewpoint-dependent coverage are absent. The 2,000-case claim is also reported without a success threshold or per-trial counts (Tables I-IV show single transforms; Table V gives means only). The central robustness claim therefore rests on an unproven guarantee and a favorable synthetic setup.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ESM-ICP, a modification of the Iterative Closest Point algorithm in which a Gaussian kernel defines per-pair weights w_i = exp(-||s_i - t_{c(i)}||^2/(2σ²)), used to build a symmetric similarity matrix M and a weighted cross-covariance matrix for SVD-based transformation estimation. The authors claim that this addresses large rotational differences and non-Gaussian noise, provide a convergence theorem, and report experiments on Stanford Bunny and ModelNet40 comparing against classical, learning-based, and KISS-Matcher baselines, with the stated finding that ESM-ICP successfully aligned over 2,000 random rotation trials within 100 iterations.","tokens_in":15106,"tokens_out":5827,"duration_ms":51929,"significance":"If the claims were supported, a training-free, simple robust-ICP variant would be a practically useful contribution for robotics and vision systems. The public implementation is a positive reproducibility feature. However, the method is a reformulation of correntropy-weighted ICP approaches already cited by the authors, the convergence theorem is not valid as stated, Equation (9) is dimensionally inconsistent, and the experimental protocol uses synthetic same-point-set data with exact correspondences. The reported tables also contain internally inconsistent metric values. These issues mean the central practical claim is not established, so the paper's significance falls well below the standard for a journal publication.","major_comments":[{"comment":"The proof asserts in Eq. (15) that E_{k+1} < E_k, but this does not follow from the stated argument. After iteration k, the correspondences c(i) and weights w_i are recomputed from the new transformation, so A_{k+1} minimizes E_k, not the reweighted objective E_{k+1}. Additionally, the proof states that the transformation updates 'satisfy ||A_{k+1} - A_k||' without giving any bound or completing the sentence, and a monotonically decreasing scalar error does not imply convergence of the transformation sequence for this nonconvex nearest-neighbor iteration. This invalidates the theoretical support for the 100%-success robustness claim made in Section V-B.","section":"Section IV-A, Theorem 1"},{"comment":"Equation (9) defines H = Σ (s_i - \\bar{s}) M (t_i - \\bar{t}) with M an N×N matrix and s_i, t_i 3D vectors. This expression is dimensionally inconsistent as written; it needs a convention (for example, treating the point vectors as rows and summing scalars, or replacing M with the scalar weight w_i) to yield the 3×3 cross-covariance matrix required by the SVD step in Eq. (10). The algorithm depends on this definition, so the method is not precisely specified.","section":"Section IV, Eq. (9)"},{"comment":"The experimental protocol generates the Source by applying a rigid transformation to the Target point cloud, guaranteeing equal cardinality and exact one-to-one correspondences. This does not exercise partial overlap, differing sampling density, or viewpoint-dependent coverage, which are the conditions that make real multi-view registration challenging. The 2,000-trial success claim is reported without a success threshold, per-trial counts, or error distributions; Tables I-IV show single examples and Table V only means. Consequently, the abstract's conclusion that ESM-ICP handles the practical limitations of current registration techniques is not supported by the evidence.","section":"Section V-B and experimental setup"},{"comment":"Several reported metric values violate the identities RMSE = sqrt(MSE) and RMSE ≥ MAE. For example, Table V Person KISS-Matcher reports MSE(t)=0.930943 but RMSE(t)=0.032311, whereas sqrt(0.930943)=0.9648; Table V Person ESM-ICP reports MSE(t)=3.17e-08 but RMSE(t)=1.07e-04, whereas sqrt(3.17e-08)=1.78e-04; Table VI Stool ESM-ICP reports MSE(t)=0.040046 but RMSE(t)=0.164954, whereas sqrt(0.040046)=0.2001. These internal inconsistencies undermine the quantitative comparisons and suggest an error in metric computation or reporting.","section":"Tables I-VI"}],"minor_comments":[{"comment":"In Algorithm 1, line 6 defines the variable d but line 7 recomputes the same distance; d is unused and should be removed or used consistently.","section":"Algorithm 1"},{"comment":"The Figure 2 caption reports the translation as (0, 0, 0.3), while the text in Section V-A states that translations are sampled uniformly in [-0.1, 0.1]; this discrepancy should be resolved.","section":"Figure 2 caption"},{"comment":"The GitHub URL in the abstract contains a space ('ESM ICP') and should be corrected to a valid hyperlink.","section":"Abstract"},{"comment":"References [20] and [22] are duplicated, and the bibliography would benefit from a careful pass to remove redundancies and align in-text citations with the numbered list.","section":"References"},{"comment":"Equation (11) states the minimization over R ∈ SO(3) only, but the objective includes τ as well; the optimization domain should be clarified as SE(3) or the dependence on τ through Eq. (10) should be made explicit.","section":"Eq. (11)"}],"recommendation":"reject","confidential_remarks":"The paper currently reads as an early draft. The convergence proof is incomplete, the central robustness claim is tested only in a same-point-set synthetic protocol, and the quantitative tables contain impossible metric values. The method is also very close to correntropy-weighted ICP formulations already present in the reference list, so the novelty claim would require stronger positioning even if the technical issues were fixed. In my view, a major revision would not be sufficient because the theoretical and empirical foundations of the main claim are both affected; substantial new experiments and a corrected analysis would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — here’s my quick read on arXiv:2505.04540. The headline: this is not a new method. Equation 7 is the standard correntropy/Gaussian weighting used in the very ICP variants the authors cite (refs [58]–[60]), and the sparse symmetric M in Eq. 8 is a notational repackaging rather than a new solver.\n\nWhat the paper does well: the weighted Procrustes step is written out cleanly, the visualization of M’s evolution toward the identity is a nice pedagogical touch, and the comparison set is broad, including a late evaluation against KISS-Matcher. The code is supposedly public, which helps.\n\nThe soft spots are serious, though. Theorem 1 is not a proof: correspondences and weights are recomputed at each iteration, so the objective in Eq. 12 changes from one step to the next; the chain E_{k+1} < E_k does not follow, and monotone decrease of a changing error would not imply global convergence even if it did. Second, the tables contain internal inconsistencies: in Table II, ICP-NL reports MAE(t) 0.8085 with RMSE(t) 0.1037, which violates RMSE ≥ MAE; Table V has MSE(t) and RMSE(t) mismatched for KISS-Matcher Person. These look like careless reporting, but they erode trust in every number. Third, the 2,000-rotation 100% success claim in V-B has no success threshold and no per-trial error distribution. Fourth, the experiments only generate Source as a rigid transform of Target, so exact one-to-one correspondences exist; no partial overlap or different sampling density, the very cases that break global registration. And σ is hand-tuned per dataset.\n\nThe paper also never acknowledges that its weighting is literally the correntropy criterion from the cited refs [58]–[60]. That is the main missing reference.\n\nVerdict: the core idea is real but not novel, and the evidence as written does not support the abstract’s claim of robust arbitrary-rotation alignment. I would still send this to peer review rather than desk-reject, because a corrected revision could be a useful within-subfield contribution. But I’d expect major revision: fix or drop the convergence theorem, correct the tables, add partial-overlap and per-trial experiments, and reposition the contribution relative to correntropy ICP. Engage, with skepticism.","headline":"A clearly written but overclaimed repackaging of correntropy-weighted ICP; the convergence proof is invalid as stated and the tables don't add up, but the idea has a real place in the registration toolkit.","tokens_in":15582,"tokens_out":4378,"would_cite":false,"duration_ms":38950,"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":"ESM-ICP replaces ICP's hard correspondences with a Gaussian exponential similarity matrix, enabling alignment under full-range rotations and non-Gaussian noise.","keywords":["point cloud registration","ICP","similarity matrix","Gaussian kernel","large rotation","non-Gaussian noise","weighted Procrustes","3D point sets"],"falsifier":"Take a real pair of scans of the same object from widely separated viewpoints, so the two clouds overlap only partially and have different sampling densities, rotate the source by roughly 180 degrees, and run ESM-ICP. If it fails to recover the alignment on cases where exact one-to-one correspondences do not exist, then the claim that the method addresses large-rotation registration in practical scenarios is not established.","tokens_in":14623,"feed_emoji":"🧩","tokens_out":10032,"duration_ms":91225,"temperature":0.7,"pith_summary":"This paper proposes ESM-ICP, a variant of the Iterative Closest Point algorithm that replaces hard, equal-weight correspondences with a Gaussian exponential similarity matrix. Each source point is matched to its closest target point, and the pair receives weight $w_i = \\exp(-\\|s_i - t_{c(i)}\\|^2/(2\\sigma^2))$, so near matches dominate and far or outlier matches are exponentially suppressed. The weighted cross-covariance is then fed through the same SVD step as classic ICP, producing an iterative soft-assignment scheme. The paper reports that this scheme aligns every one of over 2,000 rotation combinations sampled from $[-\\pi, \\pi]$ within 100 iterations and also succeeds when the source is corrupted by a clipped mixture of Gaussian noises, where standard ICP, Go-ICP, and several learning-based baselines fail. If correct, the result is a simple, training-free way to make point cloud registration reliable under the two failure modes that most often break ICP.","feed_headline":"Exponential weights let ICP align point clouds under full rotations","feed_subtitle":"A Gaussian similarity matrix suppresses bad matches, so alignment holds at rotations up to ±180° and under noise.","key_machinery":"The load-bearing object is the similarity matrix $M$, updated every iteration from the Gaussian kernel $f(x) = \\exp(-x^2/(2\\sigma^2))$. Its entry $M(i,c(i)) = w_i$ encodes the confidence that source point $s_i$ corresponds to its nearest target point $t_{c(i)}$, and symmetry makes the matrix $M = M^\\top$. This matrix converts the ordinary cross-covariance into a weighted Procrustes problem; after SVD, the rotation and translation are recovered exactly as in classic ICP. The exponential kernel carries the argument: correspondences with small residuals contribute near full weight, large residuals decay exponentially, and as the clouds align the weights tighten toward 1 so $M$ approaches the identity, smoothly handing over to standard ICP.","core_discovery":"The central claim is that a sparse, symmetric similarity matrix built from a Gaussian kernel is enough to move ICP out of the local-minima trap that large rotations create. The authors define weights from the Euclidean distance between each source point and its closest target point, build the matrix $M$ with $M(i,c(i)) = M(c(i),i) = w_i$, and compute the cross-covariance as $H = \\sum_i (s_i - \\bar{s}) M (t_i - \\bar{t})$. SVD of $H$ yields the update $R = VU^\\top$, $\\tau = -R\\bar{s} + \\bar{t}$. On ModelNet40 classes (Airplane, Person, Vase, Bottle) the reported rotation and translation errors are orders of magnitude below those of the compared geometric and learned methods, and the paper also shows $M$ evolving toward the identity matrix as alignment improves, with the weighted error $E_k$ monotonically decreasing and the algorithm reducing to classic ICP when all weights approach 1.","pith_inferences":["Inference: the 100%-success result on 2,000 synthetic rotations should not be extrapolated to partial-overlap or differently-sampled scans; a direct test on such pairs would be the natural next validation.","Inference: annealing $\\sigma$ across iterations (starting large to see broad structure, shrinking as alignment tightens) is a plausible extension that could remove the manual per-dataset tuning the paper currently requires.","Inference: because the update is a weighted Procrustes step driven by a Gaussian kernel, the method belongs to the correntropy-style family of registration algorithms; the paper's results suggest that family is a stronger competitor to global search and learned methods than previously assumed.","Inference: if extended with an explicit outlier class or a distance threshold, ESM-ICP could be dropped into LiDAR odometry and SLAM pipelines as a lightweight alternative to feature-based coarse-to-fine registration."],"forward_implications":["Pipelines that currently require a good initial guess can use ESM-ICP directly for large-rotation cases, removing the need for a separate coarse alignment stage such as feature matching or global search.","Because the method needs no training data or learned features, it can be applied to novel object classes and sensor modalities without a pretraining step.","The monotone decrease of the weighted error and the convergence of $M$ to the identity provide a built-in stopping criterion: once $M$ is nearly diagonal, further updates are essentially classic ICP.","Under non-Gaussian noise, the exponential weights suppress outlier correspondences during the covariance computation, so no explicit outlier-trimming pre-pass is required.","Since the per-iteration cost mirrors classic ICP, the approach can run at real-time rates on CPU hardware, which the paper states makes it suitable for real-time applications."],"supporting_citations":[{"why":"Supplies the SVD-based least-squares solution for rotation and translation that ESM-ICP inherits and reweights.","marker":"[1]"},{"why":"Defines classic ICP, the algorithm ESM-ICP modifies and the primary baseline in the comparison tables.","marker":"[3]"},{"why":"Point-to-plane ICP, a geometric baseline that fails under large rotations and noise in the paper's experiments.","marker":"[8]"},{"why":"Go-ICP, the globally optimal baseline that is robust in some large-rotation cases but fails in complex coupled-axis rotations and noise.","marker":"[10]"},{"why":"RPM-Net, a learning-based soft-correspondence baseline compared under large rotations.","marker":"[31]"},{"why":"Deep Closest Point, a learning-based baseline whose performance degrades when rotations exceed its limited training range.","marker":"[47]"},{"why":"PointNetLK, a learning-based baseline compared for large-rotation alignment accuracy.","marker":"[51]"},{"why":"DeepGMR, a learning-based baseline compared under transformations and noise.","marker":"[55]"},{"why":"KISS-Matcher, a recent feature-based method compared over 1,000 random transformations and outlier corruption.","marker":"[61]"}],"fun_headline_variants":["Exponential weights help ICP align clouds under full rotations","ESM-ICP: Gaussian similarity matrix beats noise and big rotations","Robust ICP via exponential similarity for point cloud registration","Exponential-based ICP outperforms on rotations and non-Gaussian noise","New matrix makes ICP robust to ±180° rotations and noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments generate the source point cloud by applying a rigid transform to the target, so every source point has an exact counterpart in the target, and the claimed success under full-range rotations is only demonstrated in that same-point-set setting with the scale parameter $\\sigma$ tuned per dataset.","fun_headline_variants_meta":{"raw":{"variants":["Exponential weights help ICP align clouds under full rotations","ESM-ICP: Gaussian similarity matrix beats noise and big rotations","Robust ICP via exponential similarity for point cloud registration","Exponential-based ICP outperforms on rotations and non-Gaussian noise","New matrix makes ICP robust to ±180° rotations and noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3351,"prompt_tokens":1009,"completion_tokens":2342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2258}},"tokens_in":625,"tokens_out":2342,"duration_ms":15900,"temperature":1.0,"reasoning_tokens":2258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:26:04.333267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real pair of scans of the same object from widely separated viewpoints, so the two clouds overlap only partially and have different sampling densities, rotate the source by roughly 180 degrees, and run ESM-ICP. If it fails to recover the alignment on cases where exact one-to-one correspondences do not exist, then the claim that the method addresses large-rotation registration in practical scenarios is not established.","supporting_citations":[{"cited_title":"Least-Squares Fitting of Two 3-D Point Sets,","cited_arxiv_id":null,"evidence_quote":"Supplies the SVD-based least-squares solution for rotation and translation that ESM-ICP inherits and reweights."},{"cited_title":"A method for registration of 3-D shapes,","cited_arxiv_id":null,"evidence_quote":"Defines classic ICP, the algorithm ESM-ICP modifies and the primary baseline in the comparison tables."},{"cited_title":"Linear Least-Squares Optimization for Point-to-Plane ICP Surface Registration,","cited_arxiv_id":null,"evidence_quote":"Point-to-plane ICP, a geometric baseline that fails under large rotations and noise in the paper's experiments."},{"cited_title":"Go-ICP: A globally optimal solution to 3D ICP point-set registration,","cited_arxiv_id":null,"evidence_quote":"Go-ICP, the globally optimal baseline that is robust in some large-rotation cases but fails in complex coupled-axis rotations and noise."},{"cited_title":"RPM-Net: Robust Point Matching Using Learned Features,“","cited_arxiv_id":null,"evidence_quote":"RPM-Net, a learning-based soft-correspondence baseline compared under large rotations."},{"cited_title":"Deep Closest Point: Learning rep- resentations for point cloud registration,","cited_arxiv_id":null,"evidence_quote":"Deep Closest Point, a learning-based baseline whose performance degrades when rotations exceed its limited training range."},{"cited_title":"PointNetLK: Robust and efficient point cloud registration using Point- Net,","cited_arxiv_id":null,"evidence_quote":"PointNetLK, a learning-based baseline compared for large-rotation alignment accuracy."},{"cited_title":"A robust algorithm for point set registration using mixture of Gaussians,","cited_arxiv_id":null,"evidence_quote":"DeepGMR, a learning-based baseline compared under transformations and noise."},{"cited_title":"KISS-Matcher: Fast and Robust Point Cloud Registration Revisited,","cited_arxiv_id":null,"evidence_quote":"KISS-Matcher, a recent feature-based method compared over 1,000 random transformations and outlier corruption."}],"review_version":1}