{"id":"d76e2239-9dd3-4c38-b434-1bcc849787cd","arxiv_id":"1909.00866","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"3D MLESAC, a maximum-likelihood variant of RANSAC, matches RANSAC precision/recall on plant point clouds and standard benchmarks while claimed to be faster, though speed is not measured.","lead":"This paper compares five matching algorithms for 3D plant scans and introduces 3D MLESAC, a statistical variant of RANSAC. The paper argues the new method is as accurate as RANSAC and faster, which would help automated plant monitoring, but the speed evidence is missing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper claims MLESAC is much faster and computationally less intense than RANSAC, but it never measures runtime; the central efficiency claim is unsupported.","rationale":"The reader's verdict is REJECT, and I agree that rejection is appropriate. My load-bearing concern differs from the reader's stated weakest assumption: the reader focuses on the unreported and untested MLESAC likelihood model in Eq. 6, while I focus on the absence of any runtime evidence for the paper's central efficiency claim. Both are valid, but the efficiency claim is more directly tied to the abstract and conclusions, and it is not supplied by fixing the likelihood model; it requires direct measurement. The reader's rationale does mention 'the central efficiency claim has no runtime evidence,' so there is partial agreement. The concrete test I propose would settle whether the efficiency claim holds. Since the paper currently lacks that evidence, the verdict remains REJECT rather than being changed to a conditional acceptance; however, if the authors release code or timings and correct the metric anomalies, re-evaluation would be appropriate.","tokens_in":8338,"tokens_out":3535,"duration_ms":37100,"concrete_test":"Add a controlled runtime benchmark on the plant dataset's 80 scenes: run the authors' MATLAB implementations of RANSAC and MLESAC under the exact Table 2 settings, separate the keypoint/descriptor stages from the correspondence-grouping stage, and report mean and standard deviation of grouping time per scene, including actual iterations executed if any early stopping occurs. If MLESAC is not faster than RANSAC at matched precision/recall settings, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline contribution is MLESAC's computational efficiency: the abstract states 'MLESAC is efficient and is computationally less intense than 3D random sample consensus (RANSAC),' and Section 5 concludes that 'in general, MLESAC is much faster compared to RANSAC.' Yet no runtime measurement, timing protocol, CPU-time table/figure, or per-iteration complexity comparison appears anywhere in Sections 4 or 5. Table 2 fixes outer iteration counts at N_ransac = N_mlesac = 1000, but this only equalizes iterations; it says nothing about the actual wall-clock cost of evaluating the likelihood in Eqs. 6-8 versus RANSAC's inlier-count evaluation. The statements 'taking less time' and 'quite faster' are qualitative assertions. This is the most load-bearing concern because the claimed practical advantage of the proposed algorithm for large plant point clouds rests specifically on computational cost, not on precision/recall alone. Even if the residual model in Eq. 6 were perfectly correct, the efficiency claim would still lack evidence without runtime data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a comparative evaluation of four 3D correspondence grouping algorithms—NNSR, RANSAC, Geometric Consistency (GC), and 3D Hough Voting (3DHV)—and introduces a 3D variant of MLESAC, formed by replacing RANSAC's inlier count with a negative log-likelihood criterion based on a Gaussian-uniform mixture model of correspondence residuals. Experiments are conducted on plant point clouds and on the B3R and U3M benchmarks under varying noise, downsampling, and overlap. Precision and recall against a ground-truth affine transformation are reported. The main conclusions are that RANSAC and MLESAC outperform 3DHV and GC, and that MLESAC is much faster than RANSAC, making it preferable for large plant point clouds.","tokens_in":8563,"tokens_out":7174,"duration_ms":68483,"significance":"The application domain is relevant: 3D correspondence grouping is a bottleneck for plant phenotyping pipelines, and the paper uses real plant scans in addition to standard benchmarks. The comparison across noise, density, and overlap is a useful experimental design. However, the central claims are not supported by the evidence. The efficiency claim for MLESAC is made without any runtime measurements. Key parameters of the MLESAC likelihood (σ, ν, γ) and the ground-truth threshold t are not reported, so the experiments are not reproducible. Moreover, recall values in Figures 3 and 4 exceed the theoretical maximum of 1, indicating a miscalculation or mislabeling of the metric. Because these issues affect the pillar conclusions of the paper, the current significance is low; the evaluation would need to be redone before the results can be trusted.","major_comments":[{"comment":"The claim that MLESAC is \"computationally less intense\" and \"much faster\" than RANSAC is never backed by any timing measurement, complexity analysis, or per-iteration cost comparison. Table 2 sets N_ransac = N_mlesac = 1000, but equalizing iteration counts does not address the cost of evaluating the likelihood in Eqs. (6)-(8) versus RANSAC's inlier counting. Without runtime data, the primary practical advantage claimed for MLESAC is unsubstantiated.","section":"Abstract; Section 5; Sections 4.2.1-4.2.2"},{"comment":"The parameters σ (Gaussian standard deviation), ν (uniform width), and γ (mixing weight) are not reported anywhere in the manuscript or Table 2. The text states that γ is estimated as in [18], but no procedure or values are given. This makes the 3D MLESAC implementation irreproducible and leaves open the possibility that its apparent performance advantage is an artifact of parameter choices. In addition, no evidence is provided that the Gaussian-uniform residual model, together with the independence assumption in Eq. (7), is appropriate for 3D plant point-cloud matching residuals.","section":"Section 3.5, Eq. (6)"},{"comment":"The correctness threshold t is never specified. Since precision and recall in Eqs. (10)-(11) and all curves in Figures 3-5 are computed against this threshold, the absolute numbers and the relative ranking of algorithms depend on an unknown quantity. Without t, the experiments cannot be reproduced or independently verified.","section":"Section 4, Eq. (9)"},{"comment":"The recall axes extend beyond 1 (to 1.1 and 1.2 in the downsampling panels). Given the definition Recall = |Ccor|/|CGT| and the fact that Ccor is the set of correctly judged correspondences, recall cannot exceed 1 unless Ccor is not actually a subset of CGT. The authors must clarify whether the plotted quantity is recall or a different measure (e.g., the number of correct inliers normalized by something else). As it stands, the reported metric is self-contradictory and undermines the quantitative comparison.","section":"Section 4, Eq. (11) and Figures 3-4"},{"comment":"The NNSR threshold t_nnsr is listed as 20. In Eq. (1), the test is a ratio of two Euclidean distances; values above 1 do not filter any matches, and a threshold of 20 is effectively non-selective. If the intended value is 0.20 or similar, the table must be corrected. Otherwise, the initial correspondence sets fed to all grouping algorithms are essentially unpruned, which would bias the entire evaluation.","section":"Table 2, first row"}],"minor_comments":[{"comment":"The acronym MLEASAC appears in the abstract and Section 3.5 heading, while MLESAC is used elsewhere; please standardize.","section":"Throughout"},{"comment":"NNSR is described as one of the four compared CG algorithms in the Introduction, but it does not appear in the result plots. Clarify whether NNSR serves only as the initial matcher or as a CG algorithm, and whether its grouping performance was evaluated.","section":"Introduction; Figures 3-5"},{"comment":"The text \"26 model of sorghum, tobacco and tomato plants\" is unclear; specify the number of plant models per species/condition and the total number of model-scene pairs.","section":"Section 4.1.1"},{"comment":"The thresholds t_ransac, t_GC, and t_mlesac are given as 0.01 without units or reference to point-cloud resolution; specify whether these are in meters, multiples of point resolution, or normalized units.","section":"Table 2"},{"comment":"The manuscript contains numerous typos (e.g., \"belongng\", \"V oting\", \"precious and recall\" in Section 4.2.3, \"UW A\" for UWA). A thorough proofreading pass is needed.","section":"Throughout"},{"comment":"The caption says \"model (right) and scene (left)\" but the figure appears to show the opposite; please verify the labeling.","section":"Figure 2 caption"}],"recommendation":"reject","confidential_remarks":"The paper is not ready for publication in its current form. The absence of any runtime measurement for an efficiency claim is a serious editorial concern, and the recall >1 issue suggests the evaluation pipeline may be flawed. I would encourage the authors to redo the experiments, report all parameters, and provide timings; a resubmission as a new manuscript could then be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's main practical claim—that 3D MLESAC is meaningfully faster than RANSAC on plant point clouds—appears nowhere in the results. There is no timing table, no per-iteration cost analysis, no wall-clock curve. What is actually new is modest: a first evaluation of correspondence grouping on plant scans, using standard algorithms, plus a direct 3D adaptation of Torr and Zisserman's 2D MLESAC. That is a reasonable thing to do, and if the evaluation were solid it would be a useful data point for phenotyping and agricultural robotics.\n\nWhat the paper does well: it uses three datasets (plant scans, B3R, U3M), covers NNSR/RANSAC/GC/3DHV, and includes visual results. The writing is straightforward.\n\nThe soft spots are not minor. First, the recall curves in Figs. 3 and 4 exceed 1.0, which is impossible under Eq. 11 (|C_cor|/|C_GT|). Either the metric is computed differently than defined, or the plots are mislabeled; either way the numbers cannot be trusted. Second, the efficiency claim is the headline of the abstract and conclusion, yet runtime is never measured. Table 2 fixes both iteration counts to 1000, which does not address the cost of evaluating the mixture likelihood versus simple inlier counting. Third, the MLESAC parameters sigma, nu, and gamma are not reported, and the ground-truth threshold t in Eq. 9 is missing, so the experiments are not reproducible. The NNSR threshold of 20 also looks like a typo.\n\nThe paper is not a waste of time—the question is practically relevant—but as written the evidence does not support the central claims. A serious referee could send it back for major revision: add runtime results, correct the recall computation, and report all thresholds and code. I would not cite it in this form.\n\nIf this crosses your desk, treat it as a well-intentioned study that is not yet reliable. The right venue is a workshop or an applied journal, not a top-tier CV venue.","headline":"The efficiency claim driving this paper is never measured, and the recall numbers violate their own definition—two load-bearing problems in an otherwise straightforward comparison.","tokens_in":9097,"tokens_out":4715,"would_cite":false,"duration_ms":46595,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"On 3D plant point clouds, MLESAC matches RANSAC in accuracy, beats it in speed, and both beat Hough voting and geometric consistency.","keywords":["3D point clouds","correspondence grouping","plant phenotyping","MLESAC","RANSAC","3D Hough voting","geometric consistency","precision and recall"],"falsifier":"Plot histograms of residual distances from ground-truth matches in the plant, B3R, and U3M scenes and fit the Gaussian-plus-uniform mixture: if the inlier residuals are heavy-tailed or the outlier residuals are not uniform, the likelihood optimum in Eq. (8) is not a principled objective. Alternatively, rerun the comparison after fitting $\\sigma$, $\\nu$, and $\\gamma$ per dataset; if RANSAC then matches or beats MLESAC in accuracy and speed, the paper's ranking does not generalise.","tokens_in":8126,"feed_emoji":"🌱","tokens_out":9328,"duration_ms":83382,"temperature":0.7,"pith_summary":"This paper asks which 3D correspondence grouping algorithm, the step that separates correct from false matches between two point clouds, works best when the clouds are plants. It compares RANSAC, geometric consistency, 3D Hough voting, and a ratio-based baseline, and adds 3D MLESAC, a variant that maximizes a likelihood of residual errors instead of counting inliers. On plant shoot scans and two standard 3D benchmarks, the paper claims that RANSAC and 3D MLESAC perform about equally well in precision and recall and are clearly better than 3D Hough voting and geometric consistency. It also claims MLESAC is usually the fastest, making it the practical choice for large, fine-structured plant point clouds.","feed_headline":"MLESAC beats RANSAC on speed at equal accuracy for 3D plant clouds","feed_subtitle":"Benchmark on plant and standard 3D scans: MLESAC matches RANSAC's accuracy and runs faster.","key_machinery":"The load-bearing mechanism is the 3D MLESAC objective: Eq. (6) models the distance between each transformed keypoint and its matched point as a mixture in which inliers follow a Gaussian of variance $\\sigma^2$ and outliers follow a uniform distribution over a range $\\nu$, mixed by a weight $\\gamma$. Eq. (7) multiplies these independent per-correspondence probabilities, and Eq. (8) turns the product into a negative log-likelihood; the transform that minimizes this quantity, rather than the transform with the largest inlier count, is chosen. That switch from counting to likelihood is what makes 3D MLESAC a new estimator in this paper and is the reason given for its speed and accuracy profile. The comparison is carried by a precision/recall protocol that labels a correspondence correct when the ground-truth transform maps the model point within a threshold $t$ of the scene point.","core_discovery":"The central claim, stated as the authors would state it, is that across the plant dataset, the B3R shape-retrieval benchmark, and the U3M registration benchmark, RANSAC and 3D MLESAC perform quite closely on precision and recall and are better than 3D Hough voting and geometric consistency, while MLESAC is usually much faster than RANSAC. The paper contributes a 3D extension of MLESAC in which RANSAC's inlier-counting objective is replaced by a negative log-likelihood objective: each correspondence residual is treated as a mixture of a Gaussian inlier term and a uniform outlier term, and the transformation minimizing that negative log-likelihood is selected. On the low-overlap U3M scenes, MLESAC has the best precision and recall of all methods. The practical conclusion is that MLESAC should be preferred for applications involving 3D point clouds of plants, especially when the clouds are large and complex.","pith_inferences":["If the speed advantage persists on scans an order of magnitude larger than the 80-scene sets used here, MLESAC could make real-time or near-real-time plant phenotyping feasible on low-power field hardware.","The comparison fixes all algorithm thresholds at the same values; tuning $\\sigma$, $\\nu$, and $\\gamma$ for each dataset might change the ranking, so the paper's conclusion is safest as a statement about default parameters.","The residual-mixture assumption is untested on plant clouds; a natural extension is to fit the mixture to known inlier/outlier residual histograms and to compare MLESAC against RANSAC under the fitted model.","Applying the same comparison to non-rigid or growing plant structures, where the ground-truth transform is not a single affine map, would test whether MLESAC's advantage generalizes beyond rigid scenes."],"forward_implications":["For 3D plant phenotyping pipelines, MLESAC should be the default correspondence grouping method when precision and recall matter and point clouds are large.","RANSAC remains a close competitor in accuracy, so applications already using RANSAC can expect similar matching quality if they switch to MLESAC, with lower computation time.","3D Hough voting and geometric consistency are not competitive on plant-shaped objects, so they are poor choices for plant recognition or part-identification workloads.","The 2D MLESAC formulation transfers to 3D point clouds directly, meaning downstream tasks like plant species recognition and point cloud registration can adopt it without further algorithmic changes."],"supporting_citations":[{"why":"Supplies RANSAC, the hypothesis-and-test paradigm and baseline that 3D MLESAC modifies.","marker":"[13]"},{"why":"Provides the 2D MLESAC likelihood-mixture objective that the paper extends to 3D.","marker":"[18]"},{"why":"Defines 3D correspondence grouping and the inlier-search problem the comparison targets.","marker":"[9]"},{"why":"Prior performance evaluation of 3D correspondence grouping algorithms that this paper extends to plant clouds.","marker":"[10]"},{"why":"Describes geometric consistency, one of the algorithms compared.","marker":"[14]"},{"why":"Introduces 3D Hough voting, another algorithm compared.","marker":"[16]"},{"why":"Defines the Lowe ratio test underlying the NNSR baseline for initial matches.","marker":"[21]"},{"why":"Supplies the 3D plant shoot architecture scans used as the plant dataset.","marker":"[25]"},{"why":"Supplies the B3R shape-retrieval benchmark used for noise and density robustness tests.","marker":"[26]"},{"why":"Supplies the U3M registration benchmark used for partial-overlap tests.","marker":"[27]"}],"fun_headline_variants":["MLESAC matches RANSAC accuracy, runs faster on 3D plant scans","3D MLESAC: best speed-accuracy tradeoff for plant point cloud matching","MLESAC wins on speed for 3D plant matching, equal to RANSAC precision","Faster 3D plant matching: MLESAC ties RANSAC, beats Hough voting","MLESAC fastest for 3D plant point clouds, matches RANSAC accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The MLESAC advantage rests on the assumption that errors of correct matches spread like a bell curve and errors of wrong matches spread evenly and independently; the paper never verifies this mixture on plant point clouds or reports the bell width, outlier spread, or mixing fraction it used.","fun_headline_variants_meta":{"raw":{"variants":["MLESAC matches RANSAC accuracy, runs faster on 3D plant scans","3D MLESAC: best speed-accuracy tradeoff for plant point cloud matching","MLESAC wins on speed for 3D plant matching, equal to RANSAC precision","Faster 3D plant matching: MLESAC ties RANSAC, beats Hough voting","MLESAC fastest for 3D plant point clouds, matches RANSAC accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3233,"prompt_tokens":896,"completion_tokens":2337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2221}},"tokens_in":512,"tokens_out":2337,"duration_ms":17423,"temperature":1.0,"reasoning_tokens":2221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:33:59.224455+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Plot histograms of residual distances from ground-truth matches in the plant, B3R, and U3M scenes and fit the Gaussian-plus-uniform mixture: if the inlier residuals are heavy-tailed or the outlier residuals are not uniform, the likelihood optimum in Eq. (8) is not a principled objective. Alternatively, rerun the comparison after fitting $\\sigma$, $\\nu$, and $\\gamma$ per dataset; if RANSAC then matches or beats MLESAC in accuracy and speed, the paper's ranking does not generalise.","supporting_citations":[{"cited_title":"Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography","cited_arxiv_id":null,"evidence_quote":"Supplies RANSAC, the hypothesis-and-test paradigm and baseline that 3D MLESAC modifies."},{"cited_title":"Mlesac: A new robust estimator with application to estimating image geometry","cited_arxiv_id":null,"evidence_quote":"Provides the 2D MLESAC likelihood-mixture objective that the paper extends to 3D."},{"cited_title":"In search of inliers: 3d corre- spondence by local and global voting","cited_arxiv_id":null,"evidence_quote":"Defines 3D correspondence grouping and the inlier-search problem the comparison targets."},{"cited_title":"Performance evaluation of 3d correspondence grouping algorithms","cited_arxiv_id":null,"evidence_quote":"Prior performance evaluation of 3D correspondence grouping algorithms that this paper extends to plant clouds."},{"cited_title":"3d free-form object recognition in range images using local surface patches","cited_arxiv_id":null,"evidence_quote":"Describes geometric consistency, one of the algorithms compared."},{"cited_title":"Object recognition in 3d scenes with occlusions and clutter by hough voting","cited_arxiv_id":null,"evidence_quote":"Introduces 3D Hough voting, another algorithm compared."},{"cited_title":"3d scans of plant shoot architectures, Jul 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D plant shoot architecture scans used as the plant dataset."},{"cited_title":"Performance evaluation of 3d keypoint detectors","cited_arxiv_id":null,"evidence_quote":"Supplies the B3R shape-retrieval benchmark used for noise and density robustness tests."},{"cited_title":"A novel representation and feature matching algorithm for automatic pairwise registration of range images.International Journal of Computer Vision, 66(1):19– 40, 2006","cited_arxiv_id":null,"evidence_quote":"Supplies the U3M registration benchmark used for partial-overlap tests."}],"review_version":1}