Grasp Type Estimation for Myoelectric Prostheses using Point Cloud Feature Learning

Federico Tombari; Ghazal Ghazaei; Kianoush Nazarpour; Nassir Navab

REVIEW 2 major objections 5 minor 20 references

Grasp Type Estimation for Myoelectric Prostheses using Point Cloud Feature Learning

T0 review · 2 major / 5 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 Adding surface normals to depth point clouds lets a PointNet estimate a prosthetic grasp type with 85.4% mean accuracy in about 0.03 seconds.

desk verdict A sensible PointNet application to prosthetic grasp choice, but the accuracy number is not trustworthy because train/test splits are by point cloud, not by object. read the letter →

arxiv 1908.02564 v1 pith:C3EKFMXE submitted 2019-08-07 cs.CV

Ghazal Ghazaei , Federico Tombari , Nassir Navab , Kianoush Nazarpour This is my paper · ORCID

classification cs.CV

keywords graspclassificationmyoelectricprosthesespointcloudNetsurfacenormalsdepthsensorsemi-autonomouscontrolRGB-D

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 argues that a prosthetic hand equipped with a small depth sensor can recognize the right grasp type for an object from a single depth snapshot, by classifying the object's point cloud with the PointNet architecture. The authors report that when surface normals are added to the 3D point coordinates, their model reaches a mean classification accuracy of 85.4% over four grasp types (pinch, tripod, palmar wrist neutral, palmar wrist pronated), and that a query runs in about 0.03 seconds, fast enough for semi-autonomous control. The claim matters because commercial prostheses have limited dexterity, and a vision-based grasp suggestion could reduce the user's control burden. The paper's core evidence is a comparison of a basic model (coordinates only) with an extended model (coordinates plus normals) on two public RGB-D object datasets.

What carries the argument

The central object is the PointNet architecture, a deep network that operates directly on unordered point sets; its defining component is a single symmetric function, max pooling, which aggregates per-point features into a global descriptor and makes the network invariant to the order of input points. The paper feeds each point cloud as an $n \times 6$ matrix of coordinates (position plus estimated surface normals), after zero-centering, unit-sphere normalization, and uniform sampling to 2048 points. The normals are computed from the eigenvectors of a local covariance matrix over $k=100$ neighbors. This machinery converts a raw depth snapshot into a compact grasp-relevant descriptor without expensive 3D grids or projections.

What would settle it

Run the same pipeline but split the data by object identity: hold out all point clouds of a set of objects for testing and train only on the remaining objects. If mean grasp accuracy on held-out objects falls well below the reported 85.4%, the network is not generalizing to new objects. A second check is to test on a novel object dataset with unseen instances and compare per-object accuracy.

Watch

Extended reading notes

Core claim

The authors claim that surface-normal augmented point clouds are sufficient for estimating grasp type from a single 2.5-D view: a PointNet classifier, trained on 2048-point samples with six features per point, assigns each object to one of four grasp classes. On their combined dataset of 3,797 point clouds, the extended model achieves 85.4% mean accuracy over five-fold cross-validation, an improvement of roughly 6 percentage points over the coordinates-only model, and does so in about 0.03 seconds per prediction. They further claim that depth data alone, without RGB, carries the shape information needed for grasp recognition, and that the human user can supervise the suggestion by accepting or refusing it.

Load-bearing premise

The evaluation assumes that randomly splitting point clouds measures real-world performance, even though multiple views of the same physical objects appear in both training and test sets, so the reported accuracy may reflect object memorization rather than generalization to new objects.

Editorial extensions

If this is right

A depth sensor mounted on a prosthetic hand could provide grasp suggestions in real time, since a single classification takes about 0.03 seconds.
Omitting RGB data removes the burden of background segmentation and lighting robustness, which the paper argues are harder in 2D images.
Adding surface normals improved accuracy by roughly 10% in some cross-validation folds, indicating that geometric shape cues beyond raw positions carry grasp-relevant information.
The four grasp classes correspond to postures a commercial myoelectric hand can execute, so the approach could be retrofitted to existing prostheses with modest computation.

Reading between the lines

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

Because the data split was random by point cloud, all views of a given object appear in both training and test sets; a leave-one-object-out split would be a harder and more realistic measure of generalisation to objects the prosthesis has not seen.
The paper's human-supervisor concept suggests a user study in which amputee participants accept or reject the suggested grasps; no such study is reported, so the practical benefit remains untested.
The same point-cloud classification pipeline could be adapted to output continuous grasp parameters (for example wrist orientation or finger aperture) rather than one of four discrete types, provided a suitably labeled dataset is available.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know two things about this one. First, it is a clean, well-scoped application of PointNet to a real prosthetic problem: given a single depth view, pick one of four grasps. Second, the headline 85.4% accuracy is probably inflated, because the split is random over point clouds, and the datasets contain many views of the same objects. Views of the same object appear on both sides of the split.

What is new is modest but real: using PointNet (rather than their earlier RGB CNN) on 2.5-D depth point clouds with surface normals, and showing about a 6-point gain from adding normals. The paper uses public data (Washington RGB-D, BigBIRD), reports five-fold CV and timing (~0.03s). The method is simple and the writing is honest about limitations.

The main problem is the evaluation protocol. Washington and BigBIRD give multiple views per object instance. An 80/10/10 split over 3797 point clouds almost certainly places different views of the same object in training and test. The network can then recognize the object instance rather than generalize to a new object—which is exactly the deployment scenario. The paper does not report an object-disjoint split or leave-one-object-out results, so we cannot bound the inflation. Also, the grasp labels are manually assigned using the authors' own prior protocol, so the ground truth is not independently validated. Those are fixable: rerun with object-disjoint splits, release labels and code. Until then, the accuracy claim should be treated as a demonstration of feasibility, not evidence of generalization.

This paper is for people working in assistive prosthetics and applied RGB-D perception. It is not a major scientific advance, but it is a reasonable engineering contribution with a clear flaw in the evaluation. A serious referee could fix it by asking for an object-disjoint evaluation. I would not cite the accuracy number, but I would mention the work as a promising direction.

Give it a serious peer review if the venue cares about applied prosthetics; the flaw is correctable, and the idea is plausible. Desk rejection would be too harsh.

Referee Report

2 major / 5 minor

Summary. The paper proposes a semi-autonomous grasp type estimation system for myoelectric prostheses that uses a single-view depth point cloud as input to a PointNet-based classifier. The authors combine objects from the Washington RGB-D and BigBIRD datasets, manually label each object category into one of four grasp types (pinch, palmar wrist neutral, tripod, palmar wrist pronated), and train PointNet on either XYZ coordinates or XYZ plus surface normals. They report a mean accuracy of 79.3% for the basic model and 85.4% for the extended model over five-fold cross-validation, with an inference time of about 0.03 seconds per query. The paper claims that depth data alone is sufficient for grasp recognition and that adding surface normals improves accuracy by about 10%.

Significance. The application domain is relevant and timely, and the use of PointNet on single-view point clouds for grasp classification is a sensible, computationally efficient choice; the reported inference time is promising for embedded prosthetic use. The observation that surface normals improve classification is a useful empirical finding. However, the evaluation protocol suffers from a serious object-instance leakage problem that invalidates the reported accuracy as a measure of generalization to unseen objects, which is the actual deployment scenario. The manuscript also does not release code or data, limiting reproducibility, and the ground-truth labeling is self-defined without external validation. If the authors re-run the evaluation with object-disjoint splits and the high accuracy persists, the contribution would be solid; as presented, the central empirical claim is not yet supported.

major comments (2)

[III. Results and Discussion] The train/validation/test split is performed over point clouds, not over object instances. Section III states that 80% of 3797 point clouds are used for training and 10% for testing, while Section II-A describes datasets with many views per object (Washington RGB-D has ~12 views per object; BigBIRD has 600 views per object). A random point-cloud-level split therefore places different views of the same physical object in both training and test sets. Since the grasp label is assigned once per object category, the network can exploit instance-specific geometry to 'recognize' objects rather than learn grasp-relevant shape features. For a prosthetic hand, the classifier must generalize to previously unseen objects, and the reported 85.4% mean accuracy does not measure that ability. The authors should repeat the evaluation with an object-disjoint split (e.g., group point clouds by object identity before splitting, or use leave-one-object-out cross-validation) and report per-object accuracy. Without such an experiment, the headline accuracy cannot be accepted as evidence for the central claim.
[II.A Dataset] The ground-truth grasp labels are manually assigned by the authors following their own prior protocol [8], [9]. The paper does not reference an external grasp taxonomy, does not report the participation of independent labelers, and provides no inter-rater reliability or validation against a standard such as the Cutkosky taxonomy. Because the classification accuracy is computed relative to these self-defined labels, the benchmark is not externally anchored and the reported numbers may not be comparable with other grasp classification studies. The authors should justify the label protocol more rigorously, ideally by adopting a recognized taxonomy or by reporting agreement statistics, so that the accuracy numbers have clear meaning outside this specific pipeline.

minor comments (5)

[Abstract] The abstract reports 'grasp classification accuracy of up to 88%' while the results section reports 85.4% as the mean accuracy of the extended model; the paper should clarify whether 88% is a single-fold result and report the full range across folds.
[II.A.2 BigBIRD] The sentence 'sample selection per available views rate in each object category in the BigBIRD is 5 times more than the Washington RGB-D dataset' is unclear and should be rephrased to describe the actual sampling procedure.
[II.C PointNet] The experimental setup does not specify several hyperparameters needed for reproducibility, such as the number of training epochs, batch size, optimizer, and any modifications made to the original PointNet architecture; these details should be provided.
[III. Results, Figure 6] The caption of Figure 6 says 'captions are ordered as T-P, where T and P represent true and predicted labels respectively'; this notation is cryptic and should be explained more fully.
[Abstract and Introduction] The claim that 'depth data provides all the necessary object shape information' is not directly tested in this work, since no RGB-based baseline is included; the authors should either add such a comparison or soften the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper is an empirical evaluation and the reported accuracy is not equivalent to its training inputs by construction.

full rationale

The paper's central claim is an empirical classification accuracy (about 79% basic, 85% extended) measured on a held-out test split of point clouds. There is no derivation chain in which a predicted quantity is algebraically identical to a fitted input. The PointNet architecture is imported from external work (Qi et al., ref [13]), and the data are external Washington RGB-D and BigBIRD datasets. The ground-truth grasp labels are manually assigned 'following the process adopted in [8], [9]', which are self-citations, but this is an annotation protocol rather than a load-bearing mathematical premise: the labels are training targets, not outputs being re-derived. The main validity concern visible in the paper is that the random 80/10/10 split at the point-cloud level (Section III) lets multiple views of the same physical object appear in both training and test sets, so the accuracy may overstate generalization to new objects. That is a test-design/validity issue, not circularity under the strict definition used here, because the test labels are not used to fit the model and the reported accuracy is not forced to equal the training accuracy by construction. No equation in the paper reduces an output to an input, and no fitted parameter is renamed as a prediction. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central empirical claim depends on the chosen architecture, the manually defined grasp categories, and the assumption that a single depth view encodes grasp-relevant shape. No new physical or mathematical entities are introduced. The only hand-tuned numerical parameter identified is the normal-estimation neighborhood size k=100.

free parameters (1)

k (surface normal estimation neighborhood) = 100
The covariance-based normal estimation uses k=100 nearest neighbors. Section II-B states this value 'provided us with desirable results', indicating a hand-tuned parameter that affects the normal features and thus the classification accuracy.

assumptions (3)

domain assumption PointNet provides an adequate permutation-invariant representation for grasp classification from point clouds.
The paper adopts PointNet [13] as the classifier without re-deriving or validating its suitability for the grasp task; the entire pipeline depends on this architectural choice.
domain assumption Manually assigned grasp labels following the protocol of [8] and [9] are the correct ground truth.
Objects are labeled into four grasp groups by the authors; no standardized taxonomy or inter-rater validation is used, so the reported accuracy is relative to a self-defined label set.
domain assumption A single 2.5-D depth view contains sufficient shape information to determine the appropriate grasp type.
The method uses one depth snapshot per object. The paper argues depth is sufficient but does not compare with multi-view or RGB baselines within the same evaluation, so the sufficiency of a single view is an unvalidated premise.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Grasp Type Estimation for Myoelectric Prostheses using Point Cloud Feature Learning." pith.science (2026). https://pith.science/paper/C3EKFMXE

@misc{pith2026190802564,
  author       = {Pith},
  title        = {Pith review of: Grasp Type Estimation for Myoelectric Prostheses using Point Cloud Feature Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3EKFMXE}},
  note         = {Machine review of arXiv:1908.02564}
}

read the original abstract

Prosthetic hands can help people with limb difference to return to their life routines. Commercial prostheses, however have several limitations in providing an acceptable dexterity. We approach these limitations by augmenting the prosthetic hands with an off-the-shelf depth sensor to enable the prosthesis to see the object's depth, record a single view (2.5-D) snapshot, and estimate an appropriate grasp type; using a deep network architecture based on 3D point clouds called PointNet. The human can act as the supervisor throughout the procedure by accepting or refusing the suggested grasp type. We achieved the grasp classification accuracy of up to 88%. Contrary to the case of the RGB data, the depth data provides all the necessary object shape information, which is required for grasp recognition. The PointNet not only enables using 3-D data in practice, but it also prevents excessive computations. Augmentation of the prosthetic hands with such a semi-autonomous system can lead to better differentiation of grasp types, less burden on user, and better performance.

Figures

Figures reproduced from arXiv: 1908.02564 by the authors.

**Figure 3.** The original PointNet unified architecture for clas [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** The accuracy curve for training of the a) basic and [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Illustration of confusion matrices for a)the basic and [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Samples of object point clouds from different views [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

[8]

An exploratory study on the use of convolutional neural networks for object grasp classiﬁcation,

G. Ghazaei, A. Alameer, P. Degenaar, G. Morgan, and K. Nazarpour, “An exploratory study on the use of convolutional neural networks for object grasp classiﬁcation,” in Intelligent Signal Processing,2nd IET International Conference on , pp. 1–5, 2015

work page 2015
[9]

Deep learning-based artiﬁcial vision for grasp classiﬁcation in myoelectric hands,

G. Ghazaei, A. Alameer, P. Degenaar, G. Morgan, and K. Nazar- pour, “Deep learning-based artiﬁcial vision for grasp classiﬁcation in myoelectric hands,” Journal of Neural Engineering , vol. 14, no. 3, p. 036025, 2017

work page 2017
[1]

Advances in control of multi-functional powered upper-limb prostheses,

K. Nazarpour, C. Cipriani, D. Farina, and T. D. Kuiken, “Advances in control of multi-functional powered upper-limb prostheses,” IEEE Transactions on Neural Systems and Rehabilitation Engineering , pp. 711–715, 2014

work page 2014
[2]

The role of feed-forward and feedback processes for closed-loop prosthesis control,

I. Saunders and S. Vijayakumar, “The role of feed-forward and feedback processes for closed-loop prosthesis control,” J Neuroeng Rehabil, vol. 8, no. 60, pp. 1–12, 2011

work page 2011
[3]

Improved prosthetic hand control with concurrent use of myoelectric and inertial measurements,

A. Krasoulis, I. Kyranou, M. S. Erden, K. Nazarpour, and S. Vi- jayakumar, “Improved prosthetic hand control with concurrent use of myoelectric and inertial measurements,” Journal of Neuroengineering and Rehabilitation , vol. 14, no. 1, p. 71, 2017

work page 2017
[4]

Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,

M. Atzori, M. Cognolato, and H. M ¨uller, “Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,” Frontiers in Neurorobotics, vol. 10, 2016

work page 2016
[5]

Cognitive vision system for control of dexterous pros- thetic hands: experimental evaluation,

S. Do ˇsen, C. Cipriani, M. Kosti ´c, M. Controzzi, M. C. Carrozza, and D. B. Popovi´c, “Cognitive vision system for control of dexterous pros- thetic hands: experimental evaluation,” Journal of Neuroengineering and Rehabilitation , vol. 7, no. 1, p. 42, 2010

work page 2010
[6]

Closed-loop control of grasping with a myoelectric hand prosthe- sis: Which are the relevant feedback variables for force control?,

A. Ninu, S. Dosen, S. Muceli, F. Rattay, H. Dietl, and D. Farina, “Closed-loop control of grasping with a myoelectric hand prosthe- sis: Which are the relevant feedback variables for force control?,” IEEE transactions on Neural Systems and Rehabilitation Engineering , vol. 22, no. 5, pp. 1041–1052, 2014

work page 2014

Show all 20 references

[7]

Stere- ovision and augmented reality for closed-loop control of grasping in hand prostheses,

M. Markovic, S. Dosen, C. Cipriani, D. Popovi ´c, and D. Farina, “Stere- ovision and augmented reality for closed-loop control of grasping in hand prostheses,” Journal of Neural Engineering , vol. 11, no. 4, p. 046001, 2014

2014
[10]

Megane Pro: myo-electricity, visual and gaze tracking integration as a resource for dexterous hand prosthetics,

F. Giordaniello, M. Cognolato, M. Graziani, A. Gijsberts, V . Gregori, G. Saetta, A.-G. M. Hager, C. Tiengo, F. Bassetto, P. Brugger, B. Caputo, H. M ¨uller, and M. Atzori, “Megane Pro: myo-electricity, visual and gaze tracking integration as a resource for dexterous hand pros...

2017
[11]

Sensor fusion and computer vision for context-aware control of a multi degree-of-freedom prosthesis,

M. Markovic, S. Dosen, D. Popovic, B. Graimann, and D. Farina, “Sensor fusion and computer vision for context-aware control of a multi degree-of-freedom prosthesis,” Journal of Neural Engineering , vol. 12, no. 6, p. 066022, 2015

2015
[12]

Convolutional networks for images, speech, and time series,

Y . LeCun and Y . Bengio, “Convolutional networks for images, speech, and time series,” The handbook of brain theory and neural networks , vol. 3361, p. 310, 1995

1995
[13]

Pointnet: Deep learning on point sets for 3d classiﬁcation and segmentation,

C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classiﬁcation and segmentation,” arXiv preprint arXiv:1612.00593, 2016

2016 arXiv
[14]

A large-scale hierarchical multi- view rgb-d object dataset,

K. Lai, L. Bo, X. Ren, and D. Fox, “A large-scale hierarchical multi- view rgb-d object dataset,” in Robotics and Automation (ICRA), 2011 IEEE International Conference on , pp. 1817–1824, IEEE, 2011

2011
[15]

Washington RGB-D dataset

“Washington RGB-D dataset.” https://rgbd-dataset.cs. washington.edu/index.html. Accessed: 2017-11-01

2017
[16]

BigBIRD ((Big) Berkeley Instance Recognition Dataset

“BigBIRD ((Big) Berkeley Instance Recognition Dataset.” http:// rll.berkeley.edu/bigbird/aliases/f186009c8c/. Ac- cessed: 2017-11-01

2017
[17]

R. B. Rusu, Semantic 3D Object Maps for Everyday Manipulation in Human Living Environments . PhD thesis, Computer Science department, Technische Universitaet Muenchen, Germany, October 2009

2009
[18]

Imagenet classiﬁcation with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classiﬁcation with deep convolutional neural networks,” in Advances in Neural Information Processing Systems , pp. 1097–1105, 2012

2012
[19]

Analysis of deep convolutional neural network archi- tectures,

J. van Doorn, “Analysis of deep convolutional neural network archi- tectures,” 2014

2014
[20]

Batch normalization: Accelerating deep network training by reducing internal covariate shift,

S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International Conference on Machine Learning , pp. 448–456, 2015

2015

Pith tools

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

[1] [8]

An exploratory study on the use of convolutional neural networks for object grasp classiﬁcation,

G. Ghazaei, A. Alameer, P. Degenaar, G. Morgan, and K. Nazarpour, “An exploratory study on the use of convolutional neural networks for object grasp classiﬁcation,” in Intelligent Signal Processing,2nd IET International Conference on , pp. 1–5, 2015

work page 2015

[2] [9]

Deep learning-based artiﬁcial vision for grasp classiﬁcation in myoelectric hands,

G. Ghazaei, A. Alameer, P. Degenaar, G. Morgan, and K. Nazar- pour, “Deep learning-based artiﬁcial vision for grasp classiﬁcation in myoelectric hands,” Journal of Neural Engineering , vol. 14, no. 3, p. 036025, 2017

work page 2017

[3] [1]

Advances in control of multi-functional powered upper-limb prostheses,

K. Nazarpour, C. Cipriani, D. Farina, and T. D. Kuiken, “Advances in control of multi-functional powered upper-limb prostheses,” IEEE Transactions on Neural Systems and Rehabilitation Engineering , pp. 711–715, 2014

work page 2014

[4] [2]

The role of feed-forward and feedback processes for closed-loop prosthesis control,

I. Saunders and S. Vijayakumar, “The role of feed-forward and feedback processes for closed-loop prosthesis control,” J Neuroeng Rehabil, vol. 8, no. 60, pp. 1–12, 2011

work page 2011

[5] [3]

Improved prosthetic hand control with concurrent use of myoelectric and inertial measurements,

A. Krasoulis, I. Kyranou, M. S. Erden, K. Nazarpour, and S. Vi- jayakumar, “Improved prosthetic hand control with concurrent use of myoelectric and inertial measurements,” Journal of Neuroengineering and Rehabilitation , vol. 14, no. 1, p. 71, 2017

work page 2017

[6] [4]

Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,

M. Atzori, M. Cognolato, and H. M ¨uller, “Deep learning with convo- lutional neural networks applied to electromyography data: A resource for the classiﬁcation of movements for prosthetic hands,” Frontiers in Neurorobotics, vol. 10, 2016

work page 2016

[7] [5]

Cognitive vision system for control of dexterous pros- thetic hands: experimental evaluation,

S. Do ˇsen, C. Cipriani, M. Kosti ´c, M. Controzzi, M. C. Carrozza, and D. B. Popovi´c, “Cognitive vision system for control of dexterous pros- thetic hands: experimental evaluation,” Journal of Neuroengineering and Rehabilitation , vol. 7, no. 1, p. 42, 2010

work page 2010

[8] [6]

Closed-loop control of grasping with a myoelectric hand prosthe- sis: Which are the relevant feedback variables for force control?,

A. Ninu, S. Dosen, S. Muceli, F. Rattay, H. Dietl, and D. Farina, “Closed-loop control of grasping with a myoelectric hand prosthe- sis: Which are the relevant feedback variables for force control?,” IEEE transactions on Neural Systems and Rehabilitation Engineering , vol. 22, no. 5, pp. 1041–1052, 2014

work page 2014