Improving analytical color and texture similarity estimation methods for dataset-agnostic person reidentification

Nikita Gabdullin

REVIEW 5 major objections 5 minor 15 references

Improving analytical color and texture similarity estimation methods for dataset-agnostic person reidentification

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that smoothing color histograms before bin compression and comparing textures through a pre-configured latent-space autoencoder improve a dataset-agnostic person re-id method to rank-1 92.9 on Market1501.

desk verdict Honest, incremental work undermined by a test-set-tuned hyperparameter search and the paper's own admission that color/texture features can't push performance further. read the letter →

arxiv 2412.05076 v2 pith:JW4FGQFW submitted 2024-12-06 cs.CV

Nikita Gabdullin This is my paper · ORCID

classification cs.CV

keywords personreidentificationanalyticalfeaturesCIE-Labcolorspacehistogramsmoothingsupervisedautoencoderlatentdistancedataset-agnosticedgecomputing

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 is trying to establish that two cheap modifications to an analytical, dataset-agnostic person-reidentification pipeline improve its accuracy without sacrificing interpretability or edge-device feasibility. First, smoothing color histograms in CIE-Lab space before compressing 256 bins to 64 removes threshold-noise artifacts in the L channel. Second, replacing local-binary-pattern texture labels with positions in a pre-configured latent space of a supervised autoencoder gives a graded texture-similarity measure rather than a hard class label. On Market1501 the best configuration reaches rank-1 92.9, rank-10 96, and mAP 25.4, a small gain over the earlier version, and the paper argues the remaining bottleneck is the lack of features for height and physique, not color or texture.

What carries the argument

Two mechanisms carry the argument. Histogram smoothing applies a 1-D uniform filter of window length $l_f$ to the 256-bin Lab histograms before they are compressed to 64 bins, suppressing parasitic peaks that appear when bin values oscillate around the threshold used for binarization; the paper finds that smoothing before compression with $l_f=11$ works best. The pre-configured latent-space supervised autoencoder (SAE) is a compact texture classifier whose encoder is trained to place each texture sample as a point inside one of five 2D clusters (uniform, horizontal lines, vertical lines, checkered, dots), so texture similarity becomes the Euclidean distance between latent points rather than a categorical label comparison. This gives a percentage-like similarity that degrades gracefully for mixed or obscured textures.

What would settle it

Train or fix the SAE as described, then run the full pipeline on a texture-rich re-id benchmark with the texture channel weight set to 0 versus 0.3; if rank-1 and mAP do not improve with the texture channel on that data, the transfer claim fails. A simpler check is to reproduce the reported Market1501 numbers without the SAE architecture or training-set size, since the paper omits both, which means the metrics cannot be independently confirmed without them.

Watch

Extended reading notes

Core claim

The central claim is that a fully interpretable, non-learned color similarity pipeline plus a small texture autoencoder trained only on generic texture images can match the ranking quality of deep re-identification models while remaining independent of any re-id dataset. The color side works by converting regions to Lab, smoothing each channel histogram with a 1-D uniform filter before bin compression, and comparing binarized histograms with an intersection-over-union-like similarity measure. The texture side uses a supervised autoencoder whose 2D latent space is pre-configured into clusters for five clothing-texture classes, so similarity between two textures is the distance between their latent-space points. The reported experiments show rank-1 92.9, rank-10 96, and mAP 25.4 on Market1501 with the best weight setting, and the paper concludes that further color and texture tuning will not help because new feature types are needed.

Load-bearing premise

The load-bearing premise is that the texture classifier network, trained only on stock photos and Stable Diffusion-generated textures, produces similarity judgments that transfer to real surveillance re-id photos and help ranking when its channel weight is 0.3; the paper gives no architecture details, dataset size, or failure analysis to support that transfer.

Editorial extensions

If this is right

The best configuration (Experiment 11) raises rank-1 by 1.9 points and mAP by 0.4 points over the base version on Market1501 while keeping the method dataset-agnostic.
Texture similarity via latent-space distance enables query-by-description: the paper demonstrates gallery searches for 'checkered upper clothes', 'white checkered upper clothes', and 'red upper clothes and black pants' without a query image.
Because the SAE is trained only on stock and Stable Diffusion-generated textures, the pipeline remains free of re-id training data, preserving its generalization claim.
The saturation result implies that further refinement of color and texture channels is not a productive direction; person height and physique features are the named next target.

Reading between the lines

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

Not tested in the paper: if the dataset-agnostic claim is right, the same fixed weights should transfer to other re-id benchmarks such as DukeMTMC or CUHK03 with little or no drop, and that is a direct experiment the paper leaves unrun.
Because the texture SAE occupies a human-readable latent space, a natural extension is to couple it with a language model that converts free-text clothing descriptions into latent-space queries automatically; the paper only hand-constructs such queries.
The conclusion that color and texture are saturated suggests adding a perspective-normalized body-shape or silhouette channel; one cheap version would be a height-to-shoulder-width ratio per parser region, which the paper does not explore.

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

The paper is a modest, honestly-written extension of the author's own dataset-agnostic re-id pipeline, combining histogram smoothing with an SAE texture channel. The best new thing in it is the text-based gallery search (Section 3.4): turning interpretable features into verbal queries is genuinely useful and not something you see in typical re-id papers. The method itself remains lightweight and training-free on re-id data, which is a real practical niche.

That said, the experimental core is too weak to support the headline claim. Every reported number comes from hand-tuning on the Market1501 test set used for the final ranking, with no validation split, no cross-validation, and no error bars. The base version (91, 96, 25) is not recomputed in the paper, so the 1.9-point rank-1 and 0.4-point mAP gains are a within-set comparison to a number taken from a different paper. The SAE texture component is never properly isolated: Table 2's 'no d' experiment removes the distance channel, not the texture channel, so the marginal contribution of the SAE is unknown. And the paper itself, in Section 3.5, concedes that Market1501 has very limited texture variability and that further color/texture feature work cannot improve overall performance. That undercuts the rationale for the proposed modifications.

I also worry about the SAE's transfer claim. The texture model is trained on stock images and Stable Diffusion output, with no architecture details, dataset size, or failure analysis for how it behaves on real surveillance crops. The paper shows a few qualitative search results, which are nice, but a few cherry-picked cases aren't evidence of dataset-agnostic texture similarity.

What's solid: the author is transparent about the limits, the method is interpretable and edge-friendly, and the discussion of what features are actually needed (height, build) is reasonable. But as a contribution, this is a parameter study of the author's own prior method, and the experimental protocol doesn't meet the standard for a published result. It's a workshop-level report at best.

I would not send this to a serious referee in its current form. The author would need to add a proper train/validation split or cross-validation, error bars, a recomputed baseline, and an ablation that isolates the SAE channel. If they can show the texture channel helps on a dataset with real texture diversity, that would be worth revisiting.

Referee Report

5 major / 5 minor

Summary. The paper proposes two modifications to an existing dataset-agnostic person re-identification pipeline: (1) smoothing of L-channel color histograms before bin compression in CIE-Lab color space, and (2) replacing LBP texture analysis with a supervised autoencoder (SAE) with a pre-configured two-dimensional latent space, trained on a custom texture dataset. The method is evaluated on Market1501, reporting rank-1 92.9, rank-10 96, and mAP 25.4 for the best configuration (Table 2, experiment 11), compared with the base version's 91, 96, and 25 taken from prior work. The paper also demonstrates text-based gallery search using interpretable features. The authors conclude that color and texture features cannot be improved further and that qualitatively new features are needed.

Significance. If the reported gains were robust, the work would be a useful incremental contribution to low-resource, dataset-agnostic re-identification, particularly for edge deployment. The interpretable, text-searchable feature representation (Section 3.4) is a genuinely interesting capability. The paper's reliance on analytical features and a re-id-agnostic texture model is a refreshing alternative to large end-to-end re-id networks. However, the empirical evidence for the central claim is weak: the reported improvements are small, derived from test-set tuning without validation, and the paper itself states that color and texture improvements cannot further improve performance on Market1501. The significance as a methodological advance is therefore not established.

major comments (5)

[Section 3.2, Table 2] The paper selects experiment 11 as the final configuration by comparing multiple weight settings directly on the Market1501 test split, then reports the same test split's metrics as the headline result. No validation split, cross-validation, or error bars are provided. The 1.9-point rank-1 and 0.4-point mAP differences between experiment 11 and experiment 1 are therefore in-sample optima and could be due to overfitting to this particular test set. This undermines the claim that the proposed modifications improve performance in a dataset-agnostic manner.
[Section 3.1, Table 1] The base version row (91, 96, 25) is not re-computed in this paper or even defined with the exact preprocessing of the modified pipeline; it is taken from Ref. [5]. Without a within-paper baseline run under identical conditions, the reader cannot determine whether the differences in Tables 1 and 2 are due to the proposed changes or to uncontrolled implementation differences. This affects all comparative claims in the paper.
[Section 2.2, Table 2] The SAE texture component is never ablated in isolation. Experiment 5 ('no d') removes the d channel while keeping the texture channel t, so it does not measure the marginal contribution of t. The claim that SAE texture similarity contributes positively to ranking is therefore not directly supported by any experiment in the paper. An ablation that varies t from 0 to a positive value while holding all other weights fixed is needed.
[Section 3.5] The paper states that 'Market1501 has very limited texture variability' and that 'further improvement of color and texture features cannot improve the overall performance of the method.' This directly contradicts the attribution of the observed rank-1 and mAP gains to the histogram-smoothing and SAE texture modifications. If the dataset has little texture variability, then the texture SAE cannot be the cause of a measurable improvement, and the paper's central mechanistic claim is left unsupported.
[Section 2.2] The SAE texture model is a central novel component, but the paper omits essential details: the architecture (encoder/decoder layers, output dimensionality), the size and composition of the custom texture dataset, the training procedure, and how the latent space is 'pre-configured' beyond referring to Ref. [12]. The transferability of this model to real surveillance re-id images is asserted rather than demonstrated, and no failure analysis or sensitivity study is provided.

minor comments (5)

[Section 2 title] The section heading 'Loss landscape analysis methodology' appears to be a copy-paste error; the section actually describes the color and texture similarity methodology.
[Section 3.1, Figure 1] Figure 1 lacks axis labels and a clear caption explaining what the threshold line represents; the captions of Figures 3-5 also do not state how the similarity score 'out of 8' or 'out of 14' is computed.
[Section 3.2] The notation in Table 2 mixes row numbers and cross-references (e.g., '9 with 6 (Table 3.1)') in a confusing way; the paper should refer to specific rows consistently.
[Section 3.4] The text-based search examples are qualitative and have no quantitative evaluation; the paper should clarify whether these are anecdotal illustrations or a systematic evaluation.
[References] The paper relies heavily on Refs. [5] and [12] for core definitions (d-channel derivation, SAE details), but the connection is not self-contained; a reader of this paper should not need to read two prior papers to understand the method.

Circularity Check

1 steps flagged · score 5.0 of 10

The reported Market1501 gain is an in-sample grid-search optimum: weights and smoothing are selected on the test split, so the improvement is a fitted quantity rather than an out-of-sample prediction.

fitted input called prediction [Section 3.1 (Table 1) and Section 3.2 (Table 2), experiment 11]
"Experiment 11. 9 with 6 (Table 3.1) 0.2 0.1 0.1 0.3 0.3 92.9 96 25.4 ... Combining these weights with the best histogram smoothing parameters obtained in Section 3.1 allows to achieve 1.9% increase in rank-1 and 0.4% increase in mAP, as shown by experiment 11 in Table 2."

Both Tables 1 and 2 are generated by evaluating every parameter combination on the same Market1501 test split that is later reported as the result. No validation split, cross-validation, or error bars are provided. Experiment 11 is the best-ranked configuration in this grid, and the claimed 1.9% rank-1 and 0.4% mAP improvements are the difference between that selected optimum and the baseline row, both measured on the same split. Thus the headline improvement is the argmax of an in-sample search, i.e., a fitted quantity, not an independent out-of-sample prediction of the proposed modifications. The final configuration is not predicted before seeing the test metrics; it is selected because it maximizes them.

full rationale

The derivation itself is not definitionally circular: color similarity is explicitly defined by Eq. (1), the combined score by Eqs. (2) and (3), and the smoothing operation is described in the text and Figure 1. The texture SAE is imported from the author's prior work [12] and the baseline numbers from [5], which is heavy self-citation, but I do not count it as circular because the cited prior work is not shown to assume the result being claimed here. The concrete circular step is in the evaluation protocol: every row in Tables 1 and 2 is scored on the same Market1501 test split, and Section 3.2 presents the best row of that grid as the achieved improvement. This matches the fitted-input-called-prediction pattern: hyperparameters are fitted to the test set, and the reported gain is the direct result of that fit. Section 3.5 further concedes that Market1501 has 'very limited texture variability' and that 'further improvement of color and texture features cannot improve the overall performance,' so the attribution of the gain to the proposed texture mechanism is not independently supported. A held-out validation split, cross-validation, or error bars would have broken the circularity; none is reported.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method depends on several components borrowed from the author's prior work without full specification: the analytical color pipeline, the SAE texture model, and the pre-configured latent space. The only new empirical contribution is a hyperparameter search on Market1501, so the free parameters are the values tuned on the test set.

free parameters (4)

Histogram smoothing filter length (lf) = 11
Best value selected from {5, 7, 9, 11, 17} using Market1501 metrics (Table 1).
Feature channel weights (wL, wa, wb, wd, wt) = (0.2, 0.1, 0.1, 0.3, 0.3)
Best combination from 11 experiments in Table 2, chosen using Market1501 test metrics.
Parser class weights (upper, pants, hair, gloves/boots, legs, other) = (8, 6, 3, 2, 1, 1)
Base weights from prior work [5]; variations in Table 3 did not improve metrics, so base is retained.
Threshold hyperparameter for distance-to-similarity conversion (d channel) = Not reported
Introduced in [5] Section 2.3.3; the value is not given in this paper.

assumptions (4)

standard math Lab color space is perceptually uniform and Euclidean distance is a valid similarity metric
Used for the d channel similarity in Section 2.1.
domain assumption Human parsing network provides accurate region masks
Parser from [10] is assumed accurate; errors lead to incorrect region comparisons, e.g., backpack parsed as upper clothes in Section 3.4.
ad hoc to paper The custom texture dataset (stock images and Stable Diffusion) is representative of real re-id textures
SAE training data is not described or released; the paper claims it preserves dataset-agnostic property.
ad hoc to paper Five texture classes (uniform, horizontal lines, vertical lines, checkered, dots) are sufficient to capture clothing texture variability
Used to configure the SAE latent space; no justification for completeness is given.

invented entities (1)

Supervised autoencoder with pre-configured latent space (SAE)
purpose: To compare textures by distances between latent-space points instead of class labels
The SAE and its latent space configuration come from the author's prior paper [12]; no architecture, training details, or external validation are provided here. Transfer to re-id is demonstrated only qualitatively and via indirect weight experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving analytical color and texture similarity estimation methods for dataset-agnostic person reidentification." pith.science (2026). https://pith.science/paper/JW4FGQFW

@misc{pith2026241205076,
  author       = {Pith},
  title        = {Pith review of: Improving analytical color and texture similarity estimation methods for dataset-agnostic person reidentification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JW4FGQFW}},
  note         = {Machine review of arXiv:2412.05076}
}

read the original abstract

This paper studies a combined person reidentification (re-id) method that uses human parsing, analytical feature extraction and similarity estimation schemes. One of its prominent features is its low computational requirements so it can be implemented on edge devices. The method allows direct comparison of specific image regions using interpretable features which consist of color and texture channels. It is proposed to analyze and compare colors in CIE-Lab color space using histogram smoothing for noise reduction. A novel pre-configured latent space (LS) supervised autoencoder (SAE) is proposed for texture analysis which encodes input textures as LS points. This allows to obtain more accurate similarity measures compared to simplistic label comparison. The proposed method also does not rely upon photos or other re-id data for training, which makes it completely re-id dataset-agnostic. The viability of the proposed method is verified by computing rank-1, rank-10, and mAP re-id metrics on Market1501 dataset. The results are comparable to those of conventional deep learning methods and the potential ways to further improve the method are discussed.

Figures

Figures reproduced from arXiv: 2412.05076 by the authors.

**Figure 1.** L channel histogram before and after smoothing; values above the threshold line are used to choose non-zero bins for binary histograms. One drawback of the proposed scheme is that it can be very sensitive to the noise inevitably present in all histograms, especially L channel ones [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Visualization of pre-configured LS of SAE with projections of two images and their [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Search results for persons wearing “checkered upper clothes” in Market1501. Num [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Search results for persons wearing “white checkered upper clothes” in Market1501. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Since multiple clothes corresponding to different parser classes are described in the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

[5]

Combining human parsing with analytical feature extraction and ranking schemes for high-generalization person reidentification,

N. Gabdullin, “Combining human parsing with analytical feature extraction and ranking schemes for high-generalization person reidentification,”Applied Sciences, vol. 13, no. 3,

work page
[12]

Latent space configuration for improved generalization in supervised autoencoder neural networks,

N. Gabdullin, “Latent space configuration for improved generalization in supervised autoencoder neural networks,” 2024. [Online]. Available: https://arxiv.org/abs/2402. 08441

work page 2024
[1]

Deep learning for person re-identification: A survey and outlook,

M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. H. Hoi, “Deep learning for person re-identification: A survey and outlook,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 6, pp. 2872–2893, 2022. 8

work page 2022
[2]

Person re-identification: A retrospective on domain specific open challenges and future trends,

A. Zahra, N. Perwaiz, M. Shahzad, and M. M. Fraz, “Person re-identification: A retrospective on domain specific open challenges and future trends,” 2022. [Online]. Available: https://arxiv.org/abs/2202.13121

arXiv 2022
[3]

Exploring generalization in deep learning,

B. Neyshabur, S. Bhojanapalli, D. McAllester, and N. Srebro, “Exploring generalization in deep learning,” 2017. [Online]. Available: https://arxiv.org/abs/1706.08947

arXiv 2017
[4]

Sensitivity and generalization in neural networks: an empirical study,

R. Novak, Y. Bahri, D. A. Abolafia, J. Pennington, and J. Sohl-Dickstein, “Sensitivity and generalization in neural networks: an empirical study,” 2018. [Online]. Available: https://arxiv.org/abs/1802.08760

arXiv 2018
[6]

Self-correction for human parsing,

P. Li, Y. Xu, Y. Wei, and Y. Yang, “Self-correction for human parsing,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 44, no. 6, pp. 3260–3271, 2022

work page 2022
[7]

Human Semantic Parsing for Person Re-identification

M. M. Kalayeh, E. Basaran, M. Gokmen, M. E. Kamasak, and M. Shah, “Human semantic parsing for person re-identification,” 2018. [Online]. Available: https://arxiv.org/abs/1804.00216

work page Pith review arXiv 2018

Show all 15 references

[8]

W. S. S. Günther Wyszecki,Color Science: Concepts and Methods, Quantitative Data and Formulae, 2nd ed. Wiley, 2000

2000
[9]

Face description with local binary patterns: Application to face recognition,

T. Ahonen, A. Hadid, and M. Pietikainen, “Face description with local binary patterns: Application to face recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 12, pp. 2037–2041, 2006

2006
[10]

Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing,

K. Gong, X. Liang, D. Zhang, X. Shen, and L. Lin, “Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing,” inIEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), July 2017

2017
[11]

Google coral-based edge computing person reidentification using human parsing combined with analytical method,

N. Gabdullin and A. Raskovalov, “Google coral-based edge computing person reidentification using human parsing combined with analytical method,” 2022. [Online]. Available: https://arxiv.org/abs/2209.11024

2022 arXiv
[13]

Scalable person re- identification: A benchmark,

L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable person re- identification: A benchmark,” in 2015 IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1116–1124

2015
[14]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” 2022. [Online]. Available: https://arxiv.org/abs/2112.10752 9

2022 arXiv
[2023]

Available: https://www.mdpi.com/2076-3417/13/3/1289

[Online]. Available: https://www.mdpi.com/2076-3417/13/3/1289

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [5]

Combining human parsing with analytical feature extraction and ranking schemes for high-generalization person reidentification,

N. Gabdullin, “Combining human parsing with analytical feature extraction and ranking schemes for high-generalization person reidentification,”Applied Sciences, vol. 13, no. 3,

work page

[2] [12]

Latent space configuration for improved generalization in supervised autoencoder neural networks,

N. Gabdullin, “Latent space configuration for improved generalization in supervised autoencoder neural networks,” 2024. [Online]. Available: https://arxiv.org/abs/2402. 08441

work page 2024

[3] [1]

Deep learning for person re-identification: A survey and outlook,

M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. H. Hoi, “Deep learning for person re-identification: A survey and outlook,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 6, pp. 2872–2893, 2022. 8

work page 2022

[4] [2]

Person re-identification: A retrospective on domain specific open challenges and future trends,

A. Zahra, N. Perwaiz, M. Shahzad, and M. M. Fraz, “Person re-identification: A retrospective on domain specific open challenges and future trends,” 2022. [Online]. Available: https://arxiv.org/abs/2202.13121

arXiv 2022

[5] [3]

Exploring generalization in deep learning,

B. Neyshabur, S. Bhojanapalli, D. McAllester, and N. Srebro, “Exploring generalization in deep learning,” 2017. [Online]. Available: https://arxiv.org/abs/1706.08947

arXiv 2017

[6] [4]

Sensitivity and generalization in neural networks: an empirical study,

R. Novak, Y. Bahri, D. A. Abolafia, J. Pennington, and J. Sohl-Dickstein, “Sensitivity and generalization in neural networks: an empirical study,” 2018. [Online]. Available: https://arxiv.org/abs/1802.08760

arXiv 2018

[7] [6]

Self-correction for human parsing,

P. Li, Y. Xu, Y. Wei, and Y. Yang, “Self-correction for human parsing,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 44, no. 6, pp. 3260–3271, 2022

work page 2022

[8] [7]

Human Semantic Parsing for Person Re-identification

M. M. Kalayeh, E. Basaran, M. Gokmen, M. E. Kamasak, and M. Shah, “Human semantic parsing for person re-identification,” 2018. [Online]. Available: https://arxiv.org/abs/1804.00216

work page Pith review arXiv 2018