Pith. sign in

REVIEW 5 major objections 6 minor 27 references

Performance comparison of 3D correspondence grouping algorithm for 3D plant point clouds

T0 review · 5 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read On 3D plant point clouds, MLESAC matches RANSAC in accuracy, beats it in speed, and both beat Hough voting and geometric consistency.

desk verdict 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. read the letter →

arxiv 1909.00866 v1 pith:2VPQYYPR submitted 2019-09-02 cs.CV

Shiva Azimi , Tapan K. Gandhi This is my paper · ORCID
classification cs.CV
keywords 3DpointcloudscorrespondencegroupingplantphenotypingMLESACRANSACHoughvotinggeometricconsistencyprecisionandrecall
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

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

  • 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.
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.

Referee Report

5 major / 6 minor

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.

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 (5)
  1. [Abstract; Section 5; Sections 4.2.1-4.2.2] 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.
  2. [Section 3.5, Eq. (6)] 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.
  3. [Section 4, Eq. (9)] 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.
  4. [Section 4, Eq. (11) and Figures 3-4] 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.
  5. [Table 2, first row] 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.
minor comments (6)
  1. [Throughout] The acronym MLEASAC appears in the abstract and Section 3.5 heading, while MLESAC is used elsewhere; please standardize.
  2. [Introduction; Figures 3-5] 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.
  3. [Section 4.1.1] 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.
  4. [Table 2] 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.
  5. [Throughout] 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.
  6. [Figure 2 caption] The caption says "model (right) and scene (left)" but the figure appears to show the opposite; please verify the labeling.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MLESAC is imported from cited external work and results are benchmarked against known ground truth.

full rationale

The paper's central algorithmic contribution is a 3D version of MLESAC, whose likelihood (Eq. 6), independence assumption (Eq. 7), and negative log-likelihood (Eq. 8) are taken from Torr and Zisserman's 2D MLESAC, explicitly cited as [18]. No parameter is fitted to the reported precision/recall curves; gamma is said to be computed per [18], and the evaluation measures precision/recall against known ground truth transformations on synthetically perturbed scenes and two external benchmarks (B3R, U3M). The claim that MLESAC is faster than RANSAC appears in the abstract and Section 4.2 without any runtime table or timing protocol, but this is an unsupported empirical assertion, not a circular derivation. The only self-reference, 'this choice is motivated by the results obtained in our previous work' for the ISS+3DSIFT detector-descriptor pair, is not load-bearing for the grouping-algorithm comparison and is not used to define any outcome. Thus no step in the derivation chain reduces to its own input.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces a new scoring procedure but no new physical or mathematical entities. All quantitative outcomes depend on hand-set thresholds, unreported likelihood parameters, synthetic scene generation, and a borrowed feature pipeline, so the central claim rests heavily on assumptions that the text does not independently support.

free parameters (5)
  • MLESAC likelihood parameters (sigma, nu, gamma) = not reported
    Equation 6 needs a Gaussian width sigma, a uniform-outlier support nu, and a mixing weight gamma; the paper cites Torr and Zisserman [18] for gamma but gives no values or implementation details.
  • Ground-truth correctness threshold t = not reported
    Equation 9 declares a correspondence correct only if |T_GT(p) - p'| <= t, and t is never specified; this threshold directly controls precision and recall numbers.
  • Distance thresholds and iteration counts = t_ransac = 0.01, t_GC = 0.01, t_mlesac = 0.01, N = 1000
    Hand-set thresholds in Table 2 are shared across methods with no sensitivity analysis and no justification.
  • ISS/3DSIFT feature settings = ~3% keypoints, non-max radius 4pr, descriptor radius 8pr, descriptor size 128
    Feature extraction settings from the authors' previous work affect every downstream algorithm but are not justified or varied here.
  • NNSR ratio threshold t_nnsr = 20
    Threshold in Eq. 1 for initial matching; chosen without justification and likely strongly affects the input correspondence set.
assumptions (4)
  • domain assumption Inlier residuals are Gaussian and outlier residuals are uniform (Eq. 6).
    Imported from 2D MLESAC without testing whether it describes plant point-cloud residuals.
  • domain assumption Residual distances d_i are independent (Eq. 7).
    The joint likelihood factorization in Eq. 7 requires independence; spatially correlated points on plant surfaces may violate it.
  • domain assumption Synthetic scenes created by rotating models and adding noise and down-sampling represent real plant scans.
    Section 4.1.1 builds scenes from the same models, so results may not transfer to real multi-view plant scans with occlusion, lighting, and scanning artifacts.
  • domain assumption ISS keypoints plus 3DSIFT descriptors are a suitable feature pipeline for plants.
    Section 4.1.4 attributes this choice to the authors' previous work without citation or comparison, so its suitability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Performance comparison of 3D correspondence grouping algorithm for 3D plant point clouds." pith.science (2026). https://pith.science/paper/2VPQYYPR

@misc{pith2026190900866,
  author       = {Pith},
  title        = {Pith review of: Performance comparison of 3D correspondence grouping algorithm for 3D plant point clouds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2VPQYYPR}},
  note         = {Machine review of arXiv:1909.00866}
}
read the original abstract

Plant Phenomics can be used to monitor the health and the growth of plants. Computer vision applications like stereo reconstruction, image retrieval, object tracking, and object recognition play an important role in imaging based plant phenotyping. This paper offers a comparative evaluation of some popular 3D correspondence grouping algorithms, motivated by the important role that they can play in tasks such as model creation, plant recognition and identifying plant parts. Another contribution of this paper is the extension of 2D maximum likelihood matching to 3D Maximum Likelihood Estimation Sample Consensus (MLEASAC). MLESAC is efficient and is computationally less intense than 3D random sample consensus (RANSAC). We test these algorithms on 3D point clouds of plants along with two standard benchmarks addressing shape retrieval and point cloud registration scenarios. The performance is evaluated in terms of precision and recall.

Figures

Figures reproduced from arXiv: 1909.00866 by the authors.

Figure 1
Figure 1. Performance comparison of 3D correspondence grouping algorithms [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization (snapshot) of sample model (right) and scene (left) from each dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance of the evaluated CG algorithms in terms of precision and recall for the plant dataset [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance of the evaluated CG algorithms in terms of precision and recall for the B3R dataset [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance of the evaluated CG algorithms in terms of precision and recall for the U3M dataset [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the evaluated CG algorithms for the three datasets [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

  1. [18]

    Mlesac: A new robust estimator with application to estimating image geometry

    Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer vision and image understanding, 78(1):138–156, 2000

    work page 2000

  2. [1]

    A meta-analysis of crop yield under climate change and adaptation

    Andrew J Challinor, J Watson, David B Lobell, SM Howden, DR Smith, and Netra Chhetri. A meta-analysis of crop yield under climate change and adaptation. Nature Climate Change, 4(4):287, 2014

    work page 2014

  3. [2]

    Lights, camera, action: high-throughput plant phenotyping is ready for a close-up

    Noah Fahlgren, Malia A Gehan, and Ivan Baxter. Lights, camera, action: high-throughput plant phenotyping is ready for a close-up. Current opinion in plant biology, 24:93–99, 2015

    work page 2015

  4. [3]

    Image analysis: the new bottleneck in plant phenotyping [applications corner]

    Massimo Minervini, Hanno Scharr, and Sotirios A Tsaftaris. Image analysis: the new bottleneck in plant phenotyping [applications corner]. IEEE signal processing magazine, 32(4):126–131, 2015

    work page 2015

  5. [4]

    Imaging system for classification of local flora of uttarakhand region

    Rachana Panwar, Kusha Goyal, Nilay Pandey, and Nitin Khanna. Imaging system for classification of local flora of uttarakhand region. In 2014 International Conference on Power, Control and Embedded Systems (ICPCES), pages 1–6. IEEE, 2014

    work page 2014

  6. [5]

    Novel low cost 3d surface model reconstruction system for plant phenotyping

    Suxing Liu, Lucia Acosta-Gamboa, Xiuzhen Huang, and Argelia Lorence. Novel low cost 3d surface model reconstruction system for plant phenotyping. Journal of Imaging, 3(3):39, 2017

    work page 2017

  7. [6]

    Local shape feature fusion for improved matching, pose estimation and 3d object recognition

    Anders G Buch, Henrik G Petersen, and Norbert Krüger. Local shape feature fusion for improved matching, pose estimation and 3d object recognition. SpringerPlus, 5(1):297, 2016

    work page 2016

  8. [7]

    Fast descriptors and correspondence propagation for robust global point cloud registration

    Huan Lei, Guang Jiang, and Long Quan. Fast descriptors and correspondence propagation for robust global point cloud registration. IEEE Transactions on Image Processing, 26(8):3614–3623, 2017

    work page 2017

Show all 27 references
  1. [8]

    Data-driven 3d voxel patterns for object category recognition

    Yu Xiang, Wongun Choi, Yuanqing Lin, and Silvio Savarese. Data-driven 3d voxel patterns for object category recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 1903–1911, 2015

    1903

  2. [9]

    In search of inliers: 3d corre- spondence by local and global voting

    Anders Glent Buch, Yang Yang, Norbert Kruger, and Henrik Gordon Petersen. In search of inliers: 3d corre- spondence by local and global voting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2067–2074, 2014

    2014

  3. [10]

    Performance evaluation of 3d correspondence grouping algorithms

    Jiaqi Yang, Ke Xian, Yang Xiao, and Zhiguo Cao. Performance evaluation of 3d correspondence grouping algorithms. In 3D Vision (3DV), 2017 International Conference on, pages 467–476. IEEE, 2017

    2017

  4. [11]

    Fast matching of binary features

    Marius Muja and David G Lowe. Fast matching of binary features. In 2012 Ninth conference on computer and robot vision, pages 404–410. IEEE, 2012

    2012

  5. [12]

    Robustifying corre- spondence based 6d object pose estimation

    Antti Hietanen, Jussi Halme, Anders Glent Buch, Jyrki Latokartano, and J-K Kämäräinen. Robustifying corre- spondence based 6d object pose estimation. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 739–745. IEEE, 2017

    2017

  6. [13]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381–395, 1981

    1981

  7. [14]

    3d free-form object recognition in range images using local surface patches

    Hui Chen and Bir Bhanu. 3d free-form object recognition in range images using local surface patches. Pattern Recognition Letters, 28(10):1252–1262, 2007

    2007

  8. [15]

    Ranking 3d feature correspondences via consistency voting

    Jiaqi Yang, Yang Xiao, Zhiguo Cao, and Weidong Yang. Ranking 3d feature correspondences via consistency voting. Pattern Recognition Letters, 117:1–8, 2019

    2019

  9. [16]

    Object recognition in 3d scenes with occlusions and clutter by hough voting

    Federico Tombari and Luigi Di Stefano. Object recognition in 3d scenes with occlusions and clutter by hough voting. In 2010 Fourth Pacific-Rim Symposium on Image and Video Technology, pages 349–355. IEEE, 2010

    2010

  10. [17]

    Rotational subgroup voting and pose clustering for robust 3d object recognition

    Anders Glent Buch, Lilita Kiforenko, and Dirk Kraft. Rotational subgroup voting and pose clustering for robust 3d object recognition. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 4137–4145. IEEE, 2017. 9 A PREPRINT - SEPTEMBER 4, 2019

    2017

  11. [19]

    Research on sift image matching based on mlesac algorithm

    Lijin Deng, Yan Piao, and Shuo Liu. Research on sift image matching based on mlesac algorithm. In Proceedings of the 2nd International Conference on Digital Signal Processing, pages 17–21. ACM, 2018

    2018

  12. [20]

    Feature correspondence via graph matching: Models and global optimization

    Lorenzo Torresani, Vladimir Kolmogorov, and Carsten Rother. Feature correspondence via graph matching: Models and global optimization. In European Conference on Computer Vision, pages 596–609. Springer, 2008

    2008

  13. [21]

    Distinctive image features from scale-invariant keypoints

    David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91–110, 2004

    2004

  14. [22]

    An efficient ransac for 3d object recognition in noisy and occluded scenes

    Chavdar Papazov and Darius Burschka. An efficient ransac for 3d object recognition in noisy and occluded scenes. In Asian Conference on Computer Vision, pages 135–148. Springer, 2010

    2010

  15. [23]

    Surface matching for object recognition in complex 3-d scenes

    A Johnson and M Hebert. Surface matching for object recognition in complex 3-d scenes. to appear in. Image and Vision Computing, 1998

    1998

  16. [24]

    Method and means for recognizing complex patterns, December 18 1962

    Paul VC Hough. Method and means for recognizing complex patterns, December 18 1962. US Patent 3,069,654

    1962

  17. [25]

    3d scans of plant shoot architectures, Jul 2017

    Saket Navlakha. 3d scans of plant shoot architectures, Jul 2017

    2017

  18. [26]

    Performance evaluation of 3d keypoint detectors

    Federico Tombari, Samuele Salti, and Luigi Di Stefano. Performance evaluation of 3d keypoint detectors. International Journal of Computer Vision, 102(1-3):198–220, 2013

    2013

  19. [27]

    A novel representation and feature matching algorithm for automatic pairwise registration of range images.International Journal of Computer Vision, 66(1):19– 40, 2006

    Ajmal S Mian, Mohammed Bennamoun, and Robyn A Owens. A novel representation and feature matching algorithm for automatic pairwise registration of range images.International Journal of Computer Vision, 66(1):19– 40, 2006. 10

    2006

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.