Estimating skeleton-based gait abnormality index by sparse deep auto-encoder

Huu Hung Huynh; Jean Meunier; Trong Nguyen Nguyen

REVIEW 4 major objections 5 minor 11 references

Estimating skeleton-based gait abnormality index by sparse deep auto-encoder

T0 review · 4 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 A gait abnormality index can be derived from single skeleton postures alone: sparse deep autoencoders trained only on normal gait score each frame by reconstruction error, and a per-axis weighted, sequence-averaged index separates normal…

desk verdict A clean, honest per-axis sparse-autoencoder gait index, but the 'best ability' claim rests on a four-subject split and a tie with the one-class SVM baseline. read the letter →

arxiv 1908.07415 v1 pith:ZDPREM25 submitted 2019-08-17 cs.CV

Trong Nguyen Nguyen , Huu Hung Huynh , Jean Meunier This is my paper · ORCID

classification cs.CV

keywords gaitabnormalityindexskeleton-basedanalysissparsedeepautoencoderreconstructionerrorone-classclassificationunsupervisedlearningdepthcamera

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 claims that a gait abnormality index can be computed directly from single skeleton postures, without hand-crafted features: three sparse deep autoencoders are trained on normal walking skeletons only, one per coordinate axis, and each new skeleton's reconstruction error becomes an abnormality score. The score for a walking sequence is a weighted average of per-frame errors, with weights set by each axis model's training error. On a dataset of nine volunteers walking on a treadmill, the per-sequence weighted index separates normal from abnormal gaits with AUC 0.945 and equal error rate 0.139, matching the best silhouette-based comparison method while using only skeletal data. If right, this would make automated gait screening unsupervised, inexpensive, and applicable to abnormal gait types never seen during training.

What carries the argument

The load-bearing object is a trio of fully connected sparse autoencoders with a 17-128-32-8-32-128-17 structure, one for each coordinate axis, trained by backpropagation on normal skeletons with a loss combining mean-square reconstruction error, a Kullback-Leibler sparsity penalty with target activation $\rho=0.05$ on the second layer, and $\ell^2$ weight decay. The mechanism is one-class reconstruction: an autoencoder trained only on normal postures reconstructs them well and abnormal postures poorly, so reconstruction error translates directly into an abnormality index. The index is the weighted sum of the three per-axis errors with weight $w_k = e^{-1}_k / \sum_k e_k$, which emphasizes axis models with lower training error; temporal context enters only as post-processing, by averaging the frame indices over a segment of 20 frames or the whole 1200-frame sequence.

What would settle it

Train the same three autoencoders on normal skeletons only, then test on gait sequences in which the pose sequence is normal but the temporal pattern is altered, for example identical poses with abnormal cadence, hesitation, or phase asymmetry, or the same subjects walking at a different treadmill speed. If the per-sequence weighted index does not separate these from normal walks (AUC near 0.5), the central claim that abnormality is contained in individual postures would be refuted.

Watch

Extended reading notes

Core claim

The paper establishes that the normality of a walking posture is largely contained in the single skeleton itself: after discarding eight joints and normalizing each coordinate axis separately to $[0,1]$, a skeleton is fed to three axis-wise deep autoencoders, and the mean squared reconstruction error is a usable per-frame abnormality index. A weighted sum of the three per-axis errors, with weights inversely related to each model's training-set error, outperforms the plain sum, and averaging the per-frame indices over a segment or the whole sequence reduces noise. The central experimental claim is that this unsupervised, feature-free pipeline outperforms a reimplemented hand-crafted skeleton-based HMM baseline and matches a reimplemented silhouette-based one-class SVM baseline, reaching a per-sequence classification error of 0.139.

Load-bearing premise

The key assumption is that abnormal gait leaves a visible mark inside a single normalized skeleton pose, so that after per-axis min-max normalization a normal-trained autoencoder's per-frame reconstruction error separates normal from abnormal movement; if the abnormality lives mainly in timing, coordination, speed, or absolute body geometry, the normalization and frame-wise scoring erase it.

Editorial extensions

If this is right

Screening for gait abnormalities no longer requires labeled abnormal examples: only normal walking skeletons are needed for training, and any deviation that changes joint coordinates becomes detectable as higher reconstruction error.
Because the index is computed per frame, the method can run online: a stream of skeleton frames yields a continually updated abnormality score, with temporal averaging used only to suppress noise.
The per-axis weighted fusion gives an explicit role to each coordinate direction; in a frontal Kinect view, sideways (X) and vertical (Y) joint motion carry more abnormality information than depth (Z).
The sparse middle-layer units act as interpretable joint-difference filters, so the same architecture that produces the index also yields a visualization of which joints and axes drive the abnormality signal.
For gait-type classification, the per-sequence weighted index (AUC 0.945, EER 0.139) supersedes the plain index sum and both reimplemented baselines on the reported dataset.

Reading between the lines

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

Because normalization scales each skeleton's per-axis coordinates to [0,1], the index is invariant to uniform body-size rescaling; a natural extension would be to test whether the same threshold holds across subjects with very different limb lengths, which the small nine-subject dataset cannot establish.
The paper's low weight on the Z-axis suggests the method depends on viewing geometry: with a side view, where most gait motion is along the optical axis, the axis weighting learned for a frontal view would likely need re-estimation.
The frame-wise scores could be fed into a temporal model such as an HMM or recurrent network instead of averaged, which might capture cadence- or coordination-based abnormalities that per-frame reconstruction error alone misses.
Thresholding individual sparse hidden-unit activations could yield a per-joint abnormality map, a clinically useful extension the paper does not pursue.

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

Colleague, quick read of Nguyen et al. The thing to know: this is a straightforward, honest application of sparse autoencoders to skeleton-based gait scoring. No hand-crafted features; three per-axis autoencoders trained only on normal skeletons; the abnormality index is a weighted sum of reconstruction errors, with weights derived from training MSE. That specific combination is not in the HMM or SVM baselines they cite, and the paper is clear that the setting is unsupervised.

Credit where due: the method is simple enough to reproduce from the text, the per-axis decomposition is sensible, and the hidden-unit visualizations give a nice interpretability check. They also reimplemented two baselines and evaluated on the same train/test split, which is more than many short papers bother to do.

The soft spots are real and concentrated in the evaluation. The test set has four subjects, each contributing nine gait sequences, so the 36 per-sequence scores are not independent; the effective sample size for subject generalization is four. There are no confidence intervals, no bootstrap, no cross-validation, and no sensitivity analysis for rho=0.05 or the segment length of 20. Table 2 shows the weighted-sum method at 0.139 per-sequence classification error, exactly tied with the one-class SVM baseline. So the Section 3.3 claim that the weighted sum 'gave the best ability of classification' is not established by the paper's own numbers — it is a tie, not a win. Since one reclassified sequence changes the rate by about 0.028, the differences between most rows are within the granularity of the test set. The AUC of 0.945 is also a single-split point estimate and could shift with a different subject partition.

There is also an explicit assumption, stated in Section 2.1 and echoed in the introduction, that abnormality is visible in per-frame normalized joint coordinates. That assumption is reasonable but untested; if gait abnormality lives mainly in temporal dynamics or body-size differences, the normalization could bury the signal. This is a legitimate concern, not a fatal one.

Overall: the core method is sound as a baseline approach and the writing is honest, but the evidence does not support superiority over simpler alternatives. The paper would need variance estimates, a plain-autoencoder ablation, hyperparameter sensitivity, and ideally released code/data to justify its stronger claims.

Who is it for? Researchers working on markerless gait assessment or reconstruction-error anomaly detection on Kinect skeletons. It is useful as a clean baseline and a clear index formula, not as a major advance.

Recommendation: if this is in the review pipeline, send it to peer review — the idea is legitimate and deserves referees, though they should demand substantial statistical revisions. If it is already published, treat it as a useful reference but do not cite the 'best ability' claim without independent validation.

Referee Report

4 major / 5 minor

Summary. The paper proposes an unsupervised gait abnormality index from Kinect skeleton data. Each skeleton is preprocessed by discarding 8 joints and min-max normalizing coordinates per axis; three sparse deep autoencoders are trained on normal walking skeletons only, one per coordinate axis. The abnormality index is a weighted sum of per-axis reconstruction errors (Eq. 2), and per-segment/per-sequence indices are obtained by averaging per-frame scores. The authors evaluate classification of normal vs. eight simulated abnormal gaits from 9 volunteers, comparing with a HMM baseline and a one-class SVM baseline. They report AUC and EER values, with the per-sequence weighted-sum index reaching AUC 0.945 and EER 0.139, and conclude that this weighted-sum index gives the best classification ability.

Significance. If the result holds, the paper offers a simple, unsupervised alternative to hand-crafted gait features, and the per-axis decomposition with sparsity constraints could aid interpretability. Strengths include a one-class formulation that matches practical settings where abnormal gaits are rare and diverse, and explicit comparison with two reimplemented baselines on a common subject split. However, the evidence supporting the central claim is currently thin: all evaluations rely on a single split with four test subjects, and the headline comparison with the one-class SVM baseline is a tie. The significance of the work would be substantially higher if the evaluation uncertainty were quantified and the claimed superiority were supported by statistically meaningful comparisons.

major comments (4)

[Section 3.3, Table 2] In Table 2, the per-sequence classification error of 'Ours (weighted sum)' is 0.139, identical to that of the One-class SVM [1]. The sentence in Section 3.3 that 'our index estimation that used a weighted sum gave the best ability of classification' is therefore not supported by the paper's own numbers; the method ties the baseline on this metric. Please provide error bars, a statistical comparison, or a clearly qualified claim.
[Sections 3.1 and 3.2, Tables 1 and 2] The evaluation uses a single subject-wise split (5 training, 4 test subjects) and reports point estimates only. The 36 per-sequence scores are not independent because they come from 4 subjects, so the effective sample size for subject generalization is 4, not 36. A single reclassified sequence changes the error rate by about 0.028, which is comparable to or larger than the differences among per-sequence rows in Table 2. I request confidence intervals, leave-one-subject-out cross-validation, or bootstrapping over subjects before the reported AUC/EER can be interpreted as evidence of superiority.
[Sections 2.1 and 2.3] The method rests on the assumption that gait abnormality is visible in per-skeleton joint positions after per-axis min-max normalization of each skeleton. The paper states this assumption but does not test it. If abnormality is largely a temporal dynamics phenomenon or is related to absolute body or joint positions, this preprocessing could remove the signal. An ablation comparing raw or body-size-normalized coordinates, or a quantitative test of the per-frame assumption, would strengthen the claim.
[Section 2.2 and Eq. (2)] Several load-bearing hyperparameters are fixed without sensitivity analysis: the sparsity target rho=0.05, the segment length of 20, and the KL/L2 penalty coefficients are not varied, and the optimization details are not reported. In addition, Eq. (2) is typeset ambiguously: as written, w_k = e^{-1}_k sum_k e_k does not define the normalized weights described in the text. Please clarify the formula and report the chosen weighting scheme.

minor comments (5)

[Section 2.2] The paper says tanh is assigned to two-thirds of the layers, but it does not specify which of the hidden layers use tanh and which use sigmoid; please state the architecture precisely.
[Figure 3] The batch-loss curves would be much easier to interpret with labeled axes and a legend; without them, the claim that the Z-model loss is less stable is difficult to verify from the figure.
[Table 2] The HMM row has no per-frame entry; please indicate why this value is unavailable or state explicitly that the method requires a sequence input.
[Section 3.3] The reimplementations of [1] and [3] are not described in enough detail to be reproduced; please provide the number of HMM states, k-means parameters, one-class SVM kernel and hyperparameters, and any tuning procedure.
[Conclusion] The final paragraph mentions Procrustes analysis as future work, but since coordinate normalization is central to the proposed preprocessing, a sensitivity experiment with an alternative normalization would be more informative than a future-work note.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the abnormality index is an unsupervised reconstruction-error score computed from label-free weights.

full rationale

The derivation chain is self-contained. Each autoencoder is trained only on normal skeletons, and the abnormality index is defined as a weighted sum of reconstruction errors on held-out skeletons. No class labels are used in training or in computing the weights of Eq. (2); those weights are derived from training-set reconstruction errors only. The claim that abnormal inputs produce high reconstruction error is an empirical assumption, not a definitional equivalence. The only self-citation is the authors' earlier HMM work [3], which appears solely as a baseline for comparison rather than as a load-bearing premise or uniqueness theorem. The numerical comparison in Table 2 is weak (a tie with the one-class SVM at 0.139 per-sequence error, based on only four test subjects), but that is a statistical-evidence concern, not circularity. No fitted parameter is renamed as a prediction, and no central result reduces to its own input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. The central result depends on manual hyperparameters, notably rho and the segment length, plus the stated assumptions that reconstruction error is a valid abnormality signal and that simulated abnormalities represent real gait pathology. Per-axis weights are formula-based from training errors and are not counted as fitted free parameters.

free parameters (3)

Sparsity target rho = 0.05
Set in Section 2.2; no sensitivity analysis is reported, and the KL penalty is central to the sparse feature claim.
Per-segment window length = 20
Used in Section 3.2 for per-segment averaging; only one length is tested despite its effect on AUC.
KL penalty and L2 regularization coefficients = not reported
The loss in Section 2.2 sums reconstruction, KL, and L2 terms without stating their relative weights, so the trained model is not fully specified.

assumptions (4)

domain assumption Normal walking skeletons form a representative one-class description, and abnormal gaits produce higher autoencoder reconstruction error.
Stated in Section 2.3; this underpins the whole index definition.
ad hoc to paper Per-skeleton, per-axis min-max normalization to [0,1] preserves the gait abnormality signal.
Section 2.1 scales every skeleton separately, which removes absolute position and scale; the paper does not test alternative normalizations.
domain assumption Simulated abnormalities such as sole padding and ankle weights are a valid proxy for real gait abnormality.
Section 3.1 defines the evaluation data; no clinical gait patients are included.
ad hoc to paper Coordinates along X, Y, and Z can be treated independently by three autoencoders.
Section 2.2 trains three independent models; the conclusion acknowledges that correlation between per-coordinate autoencoders may be worth considering.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Estimating skeleton-based gait abnormality index by sparse deep auto-encoder." pith.science (2026). https://pith.science/paper/ZDPREM25

@misc{pith2026190807415,
  author       = {Pith},
  title        = {Pith review of: Estimating skeleton-based gait abnormality index by sparse deep auto-encoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDPREM25}},
  note         = {Machine review of arXiv:1908.07415}
}

read the original abstract

This paper proposes an approach estimating a gait abnormality index based on skeletal information provided by a depth camera. Differently from related works where the extraction of hand-crafted features is required to describe gait characteristics, our method automatically performs that stage with the support of a deep auto-encoder. In order to get visually interpretable features, we embedded a constraint of sparsity into the model. Similarly to most gait-related studies, the temporal factor is also considered as a post-processing in our system. This method provided promising results when experimenting on a dataset containing nearly one hundred thousand skeleton samples.

Figures

Figures reproduced from arXiv: 1908.07415 by the authors.

**Figure 2.** The structure of our proposed auto-encoder. Each layer is fully connected to the next one. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Batch losses of 3 auto-encoders during the training stage. The loss was computed accord [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Visualization of some hidden units (that can be considered as filters) in the second layer [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 6 canonical work pages

[1]

Automatic detection of abnormal gait

Christian Bauckhage, John K Tsotsos, and Frank E Bunn. Automatic detection of abnormal gait. Image and Vision Computing, 27 0 (1): 0 108--115, 2009

work page 2009
[2]

Stanford university cs231n: Convolutional neural networks for visual recognition

Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Stanford university cs231n: Convolutional neural networks for visual recognition. Neural networks, 1, 2016

work page 2016
[3]

Skeleton-based abnormal gait detection

Trong-Nguyen Nguyen, Huu-Hung Huynh, and Jean Meunier. Skeleton-based abnormal gait detection. Sensors, 16 0 (11): 0 1792, 2016. ISSN 1424-8220. doi:10.3390/s16111792. URL http://www.mdpi.com/1424-8220/16/11/1792

doi:10.3390/s16111792 2016
[4]

Online quality assessment of human movement from skeleton data

Adeline Paiement, Lili Tao, Sion Hannuna, Massimo Camplani, Dima Damen, and Majid Mirmehdi. Online quality assessment of human movement from skeleton data. In Proceedings of British Machine Vision Conference, 2014

2014
[5]

Shotton, A

J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman, and A. Blake. Real-time human pose recognition in parts from single depth images. In CVPR 2011, pp.\ 1297--1304, June 2011. doi:10.1109/CVPR.2011.5995316

work page arXiv 2011
[6]

Shotton, R

J. Shotton, R. Girshick, A. Fitzgibbon, T. Sharp, M. Cook, M. Finocchio, R. Moore, P. Kohli, A. Criminisi, A. Kipman, and A. Blake. Efficient human pose estimation from single depth images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 0 (12): 0 2821--2840, Dec 2013. ISSN 0162-8828. doi:10.1109/TPAMI.2012.241

work page doi:10.1109/tpami.2012.241 2013
[7]

A comparative study of pose representation and dynamics modelling for online motion quality assessment

Lili Tao, Adeline Paiement, Dima Damen, Majid Mirmehdi, Sion Hannuna, Massimo Camplani, Tilo Burghardt, and Ian Craddock. A comparative study of pose representation and dynamics modelling for online motion quality assessment. Computer Vision and Image Understanding, 148: 0 136--152, 2016

work page 2016
[8]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

Show all 11 references

[9]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...
[10]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...
[11]

J. \, S. Bell

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

2017

Pith tools

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

[1] [1]

Automatic detection of abnormal gait

Christian Bauckhage, John K Tsotsos, and Frank E Bunn. Automatic detection of abnormal gait. Image and Vision Computing, 27 0 (1): 0 108--115, 2009

work page 2009

[2] [2]

Stanford university cs231n: Convolutional neural networks for visual recognition

Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Stanford university cs231n: Convolutional neural networks for visual recognition. Neural networks, 1, 2016

work page 2016

[3] [3]

Skeleton-based abnormal gait detection

Trong-Nguyen Nguyen, Huu-Hung Huynh, and Jean Meunier. Skeleton-based abnormal gait detection. Sensors, 16 0 (11): 0 1792, 2016. ISSN 1424-8220. doi:10.3390/s16111792. URL http://www.mdpi.com/1424-8220/16/11/1792

doi:10.3390/s16111792 2016

[4] [4]

Online quality assessment of human movement from skeleton data

Adeline Paiement, Lili Tao, Sion Hannuna, Massimo Camplani, Dima Damen, and Majid Mirmehdi. Online quality assessment of human movement from skeleton data. In Proceedings of British Machine Vision Conference, 2014

2014

[5] [5]

Shotton, A

J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman, and A. Blake. Real-time human pose recognition in parts from single depth images. In CVPR 2011, pp.\ 1297--1304, June 2011. doi:10.1109/CVPR.2011.5995316

work page arXiv 2011

[6] [6]

Shotton, R

J. Shotton, R. Girshick, A. Fitzgibbon, T. Sharp, M. Cook, M. Finocchio, R. Moore, P. Kohli, A. Criminisi, A. Kipman, and A. Blake. Efficient human pose estimation from single depth images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 0 (12): 0 2821--2840, Dec 2013. ISSN 0162-8828. doi:10.1109/TPAMI.2012.241

work page doi:10.1109/tpami.2012.241 2013

[7] [7]

A comparative study of pose representation and dynamics modelling for online motion quality assessment

Lili Tao, Adeline Paiement, Dima Damen, Majid Mirmehdi, Sion Hannuna, Massimo Camplani, Tilo Burghardt, and Ian Craddock. A comparative study of pose representation and dynamics modelling for online motion quality assessment. Computer Vision and Image Understanding, 148: 0 136--152, 2016

work page 2016

[8] [8]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...