Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Accuracy and Fairness of Facial Recognition Technology in Low-Quality Police Images: An Experiment With Synthetic Faces

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Under controlled synthetic degradations, facial recognition false negatives rise steeply with blur and low resolution, and error rates are highest for Black women, while false positives peak at near-baseline image quality.

desk verdict A useful synthetic experiment mapping how five degradations shift FPR/FNR by demographic group, but the Discussion misstates the paper's own FPR result and the forensic comparison is not supported by the experiment. read the letter →

arxiv 2505.14320 v1 pith:UNWAQGSG submitted 2025-05-20 cs.CV stat.AP

Maria Cuellar , Hon Kiu (James) To , Arush Mehrotra This is my paper · ORCID
classification cs.CVstat.AP
keywords accuracyfairnessimagesusedaffectedblackblurevaluate
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

Police often feed low-quality photos, such as blurry surveillance stills, into facial recognition software. This paper builds a controlled experiment to see how image quality changes the software's mistakes. The authors generate thousands of synthetic faces with StyleGAN3, label each face as male or female and as White, Black, or Asian using a separate classifier called FairFace, then degrade copies of each face in five ways: lower contrast, brighter or darker exposure, motion blur, rotated pose, and reduced resolution. Each degraded face is searched against a small database of 167 faces in a 1 to 167 identification task using DeepFace with ArcFace, an open-source recognition model.

The main findings are that false positives, claiming a wrong match, are most common when the image is close to clean, while false negatives, failing to recognize a person who is in the database, climb steadily as blur and resolution get worse. Across nearly all conditions, women and Black faces, especially Black women, have higher error rates than White men. The authors also argue that even the worst error rates they measure stay below error rates reported for fingerprint and firearm comparison in other studies.

The results come with a major caveat. All faces are synthetic, labels come from an automated classifier, and the experiment uses one recognition model with a fixed threshold. Whether the same patterns hold for real police photos of actual people is not tested here.

Extended reading notes

Core claim

Section 5 states: 'Our experiment finds that facial recognition technology (FRT) performance degrades under poor image conditions, particularly with blur, pose variation, and reduced resolution, and that this degradation is not evenly distributed across demographic groups. False positive and false negative rates increase with image degradation, disproportionately affecting individuals from marginalized race and gender groups.' If correct, this means FRT error rates are quality-dependent and demographic-disparate in the measured synthetic 1:n task.

Load-bearing premise

The transfer from this experiment to real policing depends on StyleGAN3 synthetic faces and FairFace labels being adequate proxies for real photos of actual suspects, including the same demographic error structure. The Discussion states 'we assume the data generation and labeling algorithms have no errors' and concedes the generated data may reflect a narrow type of face. If the synthetic distribution, or FairFace's label errors, differs from real police images, the headline error-rate curves and race/gender disparities could change.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper reports a controlled experiment on how five image degradations (contrast, brightness, motion blur, pose, resolution) affect the accuracy and fairness of a DeepFace/ArcFace 1:n identification system, using 50,000 StyleGAN3 synthetic faces labeled by FairFace. The authors construct a 167-face control database representative of the US census, generate treated images, and compute false positive and false negative rates overall and by race/gender. The central empirical finding, as stated in the Abstract and Section 4, is that false positives peak near baseline quality and decline with degradation, while false negatives increase with degradation, especially for blur and reduced resolution; error rates are generally higher for women and Black individuals. The Discussion, however, contains a sentence claiming both FPR and FNR increase with degradation, which directly contradicts the paper's own figures and text. The paper also includes a comparison of FRT error rates to traditional forensic disciplines and argues that FRT, if properly validated and regulated, remains a valuable investigative tool.

Significance. If the empirical curves are correct, the paper provides a useful experimental template for isolating the effects of specific image-quality factors on FRT accuracy and fairness, and it highlights a counterintuitive pattern (FPR highest at high quality) that is relevant for police practice. The use of synthetic faces allows controlled, gradual degradation at scale, and the resampling-based confidence intervals are a methodological strength. However, the paper's significance is bounded by two threats: the synthetic data and FairFace labels are treated as ground truth, and the Discussion's summary sentence misstates the reported FPR trend. Neither threat invalidates the experimental design, but the first limits external validity and the second must be corrected before the paper can be accepted.

major comments (4)
  1. [Section 5, first paragraph] The sentence "False positive and false negative rates increase with image degradation" is directly contradicted by the paper's own results in Section 4.1 and Figure 6, which show FPR peaking near baseline and declining as degradation intensifies (the "inverted U-shaped pattern"), and by Section 4.2, which states that FPR disparities "grow less pronounced as degradation intensifies." The Abstract and Section 4.1 correctly describe the FPR trend. Because this Discussion sentence is part of the paper's headline summary and appears verbatim in the reader-facing description, it is a load-bearing internal inconsistency that must be corrected to match the empirical curves.
  2. [Section A.3 (Pose)] The description of the pose baseline correction is incomplete: the sentence "This shift w" ends mid-word. The reader cannot determine how the pose curve was shifted relative to the other degradation factors, which is necessary to interpret Figure 6 and Figure 7 and to reproduce the analysis. Please provide the full correction formula and the numerical value of the shift.
  3. [Section 3.2 (Replications)] The paper states "We perform a total of 2^8 = 256 replications" but does not explain how the sample of 167 control and 167 treated faces is redrawn in each replication, nor how the empirical confidence intervals are constructed from these replications. This is central to the uncertainty estimates in Figures 6 and 7. Please clarify the resampling protocol (e.g., independent draws from the 50,000-face population for each replication, or bootstrap over a fixed sample) and report the exact procedure used to obtain the 95% intervals.
  4. [Sections 2.3 and 5 (External validity)] The paper relies on FairFace labels as ground truth for race/gender and, as stated in the Discussion, "assume[s] the data generation and labeling algorithms have no errors" while also acknowledging that the generated data "could lead to having the data show a narrow type of face." These are not merely minor caveats: the headline claims about "individuals from marginalized race and gender groups" and the comparison to real police images depend on the synthetic faces and labels being adequate proxies for real-world low-quality images. The paper should either explicitly restrict the conclusions to the synthetic distribution or provide additional evidence (e.g., a comparison with a real low-quality face dataset) that the demographic error structure transfers.
minor comments (4)
  1. [Figure captions and Section 5] The word "Degredation" in the captions of Figures 6 and 7 should be "Degradation," and "eggregious" in Section 5 should be "egregious."
  2. [Section 3.2] The expression "2 8" should be typeset as "2^8" to avoid ambiguity.
  3. [References] Reference [56] duplicates Reference [11] (both are FRVT Part 3 demographic effects); please merge or differentiate them.
  4. [Section 5] The sentence "While this decision provides a conservative estimate of system performance under real-world conditions" references which decision (likely the enforce_detection=False setting) but the antecedent is unclear; please name the setting explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-level circularity: error-rate curves are measured outputs, not fitted inputs; the two self-citations are framing-only and non-load-bearing.

full rationale

The paper's central quantities, FPR and FNR, are defined by the standard formulas in Eqs. (1)-(2) and computed from observed match outcomes of Deepface/ArcFace in a fixed 1:167 identification task; the degradation treatments (contrast, brightness, motion blur, pose, resolution) are exogenous image edits, and the demographic groupings come from FairFace labels. No parameter in this pipeline is fitted to the reported curves and then renamed as a prediction, and no theoretical uniqueness claim is imported to force a choice. The only self-citations ([26], [30]) appear in framing passages about forensic-reform lessons and about methodological criticism of prior firearm-comparison studies; they do not carry any of the experimental results, so they are not load-bearing. The Discussion does contain an internal-consistency problem that is not circularity: Section 5 says 'False positive and false negative rates increase with image degradation,' whereas Section 4.1 reports an 'inverted U-shaped pattern' with degradation acting 'as a kind of natural filter' and Section 4.2 reports FPR disparities 'grow less pronounced as degradation intensifies'; this overstatement is a correctness issue in the summary, not a reduction of output to input. The authors' own limitations (assumption that generation and labeling have no errors; possibility of a 'narrow type of face') are external-validity and measurement-error caveats, not circularity.

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

The central claim is an empirical measurement, so the ledger is dominated by domain assumptions about the synthetic pipeline rather than fitted constants. The free parameters are the operating threshold, the GAN truncation, the pose correction offset, and the degradation grids. No new entities are introduced.

free parameters (4)
  • ArcFace cosine distance threshold T = Not reported in the paper; DeepFace developer-recommended default was used.
    All FPR and FNR rates are conditional on this threshold (Section 3.2). Changing T moves the operating point and could change which demographic groups appear most error-prone.
  • StyleGAN3 truncation parameter psi = 0.7 (fixed in Section 2.1).
    Chosen to balance facial fidelity and variety. It alters the distribution of generated faces, including demographic mix and identity distinctiveness, and thereby all error rates.
  • Pose baseline correction offset = Difference between pose psi = 0 error rates and the common baseline error rate, computed from the experiment's own data.
    Appendix A.3: because pose psi = 0 images are still slightly altered, a fixed value is added to every pose error rate. This data-derived shift is applied for 'a more honest comparison' and shapes the pose curves.
  • Degradation parameter grids = Contrast alpha from 0.25 to 4; brightness beta from 0 to 100; blur strength s from 0 to 100; pose psi from -5 to 5…
    Appendix A lists the specific levels. The reported sensitivity of each degradation factor is relative to these chosen grids; narrower or wider grids could change the ranking of factors.
assumptions (5)
  • domain assumption StyleGAN3 faces trained on FFHQ are an adequate proxy for real human faces encountered in police images.
    Entire study uses synthetic faces (Section 2.1); the authors concede in Discussion that the data may show a narrow type of face and that future work should replicate with real data.
  • domain assumption FairFace predictions are correct ground-truth labels for race and gender.
    All fairness comparisons group faces by FairFace labels (Section 2.3); Discussion states 'we assume the data generation and labeling algorithms have no errors'.
  • domain assumption DeepFace with ArcFace at a fixed developer-recommended threshold represents law enforcement FRT use.
    Sections 3.1 and 3.2 choose DeepFace and ArcFace with a fixed threshold T; operational systems such as Clearview, Rekognition, and Oosto may differ in model, threshold, and interface.
  • domain assumption Latent-space pose editing with restyle e4e produces pose variation comparable to real-world non-frontal photos.
    Appendix A.3 edits pose along a precomputed direction; no validation against real pose distributions is provided.
  • standard math Empirical resampling from the synthetic population approximates the sampling distribution of error rates.
    Section 2.2 and Figure 5 use repeated resampling to build 95% confidence intervals; this is standard bootstrap-style inference and does not by itself validate the population.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accuracy and Fairness of Facial Recognition Technology in Low-Quality Police Images: An Experiment With Synthetic Faces." pith.science (2026). https://pith.science/paper/UNWAQGSG

@misc{pith2026250514320,
  author       = {Pith},
  title        = {Pith review of: Accuracy and Fairness of Facial Recognition Technology in Low-Quality Police Images: An Experiment With Synthetic Faces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNWAQGSG}},
  note         = {Machine review of arXiv:2505.14320}
}
read the original abstract

Facial recognition technology (FRT) is increasingly used in criminal investigations, yet most evaluations of its accuracy rely on high-quality images, unlike those often encountered by law enforcement. This study examines how five common forms of image degradation--contrast, brightness, motion blur, pose shift, and resolution--affect FRT accuracy and fairness across demographic groups. Using synthetic faces generated by StyleGAN3 and labeled with FairFace, we simulate degraded images and evaluate performance using Deepface with ArcFace loss in 1:n identification tasks. We perform an experiment and find that false positive rates peak near baseline image quality, while false negatives increase as degradation intensifies--especially with blur and low resolution. Error rates are consistently higher for women and Black individuals, with Black females most affected. These disparities raise concerns about fairness and reliability when FRT is used in real-world investigative contexts. Nevertheless, even under the most challenging conditions and for the most affected subgroups, FRT accuracy remains substantially higher than that of many traditional forensic methods. This suggests that, if appropriately validated and regulated, FRT should be considered a valuable investigative tool. However, algorithmic accuracy alone is not sufficient: we must also evaluate how FRT is used in practice, including user-driven data manipulation. Such cases underscore the need for transparency and oversight in FRT deployment to ensure both fairness and forensic validity.

Figures

Figures reproduced from arXiv: 2505.14320 by the authors.

Figure 1
Figure 1. Images released by the Philadelphia Police Department in public ap [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Data generation procedure. We generate a large population of synthetic [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Examples of synthetic faces generated with StyleGAN3 and labeled with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Facial images that have been modified according to parameters (brightness, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Our method to generate the conditional error rate curves. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Accuracy curves. False positive and false negative rates for different image [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Fairness curves, i.e., differential accuracy based on demographic factors: [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Lightweight Ensemble-Based Face Image Quality Assessment Method with Correlation-Aware Loss

    cs.CV 2025-09 conditional novelty 3.0 of 10

    An ensemble of MobileNetV3-Small and ShuffleNetV2 with a correlation-aware loss and test-time augmentation reaches SRCC 0.9829 and PLCC 0.9894 on the VQualA FIQA validation set.

Reference graph

Works this paper leans on

58 extracted references · 55 canonical work pages · cited by 1 Pith paper

  1. [1]

    A forensic without the science: face recognition in US criminal inves- tigations

    Garvie C. A forensic without the science: face recognition in US criminal inves- tigations. Center on Privacy & Technology at Georgetown Law. 2022;6

    work page 2022

  2. [2]

    Linden L, editor

    Clearview AI. Linden L, editor. Consecutive NIST Tests Confirm Su- periority of Clearview AI’s Facial Recognition Platform. Online; 2021. https://www.clearview.ai/press-room/consecutive-nist-tests- confirm-superiority-of-clearview-ais-facial-recognition-platform

    work page 2021

  3. [3]

    Face Recognition Vendor Test (FR VT); 2019.https:// www.nist.gov/programs-projects/face-recognition-vendor-test-frvt

    Grother P. Face Recognition Vendor Test (FR VT); 2019.https:// www.nist.gov/programs-projects/face-recognition-vendor-test-frvt

    work page 2019

  4. [4]

    MacMillan D, Ovalle D, Schaffer A. Arrested by AI: Police ig- nore standards after facial recognition matches; 2025.https: //www.washingtonpost.com/business/interactive/2025/police- artificial-intelligence-facial-recognition/?itid=sr 0 8204cd4d- 7801-4d64-986f-b99edd2636a1 03abe5cd-4d5a-4f6b-91bf-6a1a2a070092

    work page 2025

  5. [5]

    Sanford A. Artificial Intelligence Is Putting Innocent People at Risk of Being Incarcerated; 2024.https://innocenceproject.org/news/ artificial-intelligence-is-putting-innocent-people-at-risk-of- being-incarcerated/

    work page 2024

  6. [6]

    Learning to under- stand image blur

    Zhang S, Shen X, Lin Z, Mˇ ech R, Costeira JP, Moura JM. Learning to under- stand image blur. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2018. p. 6586-95

    work page 2018

  7. [7]

    Image partial blur detection and classification

    Liu R, Li Z, Jia J. Image partial blur detection and classification. In: 2008 IEEE conference on computer vision and pattern recognition. IEEE; 2008. p. 1-8

    work page 2008

  8. [8]

    Face recognition across non-uniform motion blur, illumination, and pose

    Punnappurath A, Rajagopalan AN, Taheri S, Chellappa R, Seetharaman G. Face recognition across non-uniform motion blur, illumination, and pose. IEEE Transactions on image processing. 2015;24(7):2067-82. 21

    work page 2015

Show all 58 references
  1. [9]

    Prioritising quality: investigating the influence of image quality on forensic facial comparison

    Bacci N, Briers N, Steyn M. Prioritising quality: investigating the influence of image quality on forensic facial comparison. International Journal of Legal Medicine. 2024:1-14

    2024

  2. [10]

    Pixel-Level Face Image Quality Assessment for Explainable Face Recognition

    Terh¨ orst P, et al. Pixel-Level Face Image Quality Assessment for Explainable Face Recognition. ResearchGate. 2023. Available from:https://www.researchgate.net/publication/369678929 Pixel- Level Face Image Quality Assessment for Explainable Face Recognition

    2023

  3. [11]

    Face recognition vendor test (fvrt): Part 3, demographic effects

    Grother P, Ngan M, Hanaoka K. Face recognition vendor test (fvrt): Part 3, demographic effects. National Institute of Standards and Technology Gaithers- burg, MD; 2019

    2019

  4. [12]

    Accuracy comparison across face recognition algorithms: Where are we on measuring race bias? IEEE transactions on biometrics, behavior, and identity science

    Cavazos JG, Phillips PJ, Castillo CD, O’Toole AJ. Accuracy comparison across face recognition algorithms: Where are we on measuring race bias? IEEE transactions on biometrics, behavior, and identity science. 2020;3(1):101-11

    2020

  5. [13]

    Benchmarking Algorithmic Bias in Face Recognition: An Experimental Approach Using Synthetic Faces and Human Evaluation

    Liang H, Perona P, Balakrishnan G. Benchmarking Algorithmic Bias in Face Recognition: An Experimental Approach Using Synthetic Faces and Human Evaluation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision; 2023. p. 4977-87

    2023

  6. [14]

    Learning Face Image Qual- ity from Human Assessments

    Best-Rowden L, Jain AK. Learning Face Image Qual- ity from Human Assessments. IEEE Transactions on Infor- mation Forensics and Security. 2018;13(12):3064-77. Avail- able from:https://biometrics.cse.msu.edu/Publications/Face/ BestRowdenJain FaceQualityHumanAssessments TIFS2018.pdf

    2018

  7. [15]

    The effect of broad and specific demo- graphic homogeneity on the imposter distributions and false match rates in face recognition algorithm performance

    Howard JJ, Sirotin YB, Vemury AR. The effect of broad and specific demo- graphic homogeneity on the imposter distributions and false match rates in face recognition algorithm performance. In: 2019 ieee 10th international conference on biometrics theory, applications and system...

    2019

  8. [16]

    Quantifying the extent to which race and gender features determine identity in commercial face recognition algorithms

    Howard JJ, Sirotin YB, Tipton JL, Vemury AR. Quantifying the extent to which race and gender features determine identity in commercial face recognition algorithms. arXiv preprint arXiv:201007979. 2020

    2020

  9. [17]

    Disparate impact in facial recognition stems from the broad homogeneity effect: A case study and method to resolve

    Howard JJ, Laird EJ, Sirotin YB. Disparate impact in facial recognition stems from the broad homogeneity effect: A case study and method to resolve. In: International Conference on Pattern Recognition. Springer; 2022. p. 448-64. 22

    2022

  10. [18]

    Evaluat- ing proposed fairness models for face recognition algorithms

    Howard JJ, Laird EJ, Rubin RE, Sirotin YB, Tipton JL, Vemury AR. Evaluat- ing proposed fairness models for face recognition algorithms. In: International Conference on Pattern Recognition. Springer; 2022. p. 431-47

    2022

  11. [19]

    Reformer TM, editor; 2025.https://minnesotareformer.com/2025/ 02/04/facial-recognition-in-policing-is-getting-state-by-state- guardrails/

    Gross P. Reformer TM, editor; 2025.https://minnesotareformer.com/2025/ 02/04/facial-recognition-in-policing-is-getting-state-by-state- guardrails/

    2025

  12. [20]

    Challenging facial recognition software in criminal court

    Jackson K. Challenging facial recognition software in criminal court. The Cham- pion. 2019:14-26

    2019

  13. [21]

    State of New Jersey vs

    Superior Court of New Jersey Appellate Division. State of New Jersey vs. Francisco Arteaga. Docket No. A-3078-21; 2023

    2023

  14. [22]

    NA, editor

    ACLU. NA, editor. New Jersey Appellate Division One of First Courts in Country to Rule on Constitutional Rights Related to Facial Recognition Technologies. NA; 2023. https://www.aclu-nj.org/en/press-releases/new-jersey- appellate-division-one-first-courts-country-rule-constitu...

    2023

  15. [23]

    Confronting the biased algorithm: the danger of admitting fa- cial recognition technology results in the courtroom

    Haddad GM. Confronting the biased algorithm: the danger of admitting fa- cial recognition technology results in the courtroom. Vand J Ent & Tech L. 2020;23:891

    2020

  16. [24]

    Face-to-face with facial recognition evidence: Admissibility under the post-crawford confrontation clause

    Celentino JC. Face-to-face with facial recognition evidence: Admissibility under the post-crawford confrontation clause. Mich L Rev. 2015;114:1317

    2015

  17. [25]

    NA, editor

    PCAST Working Group. NA, editor. Forensic science in criminal courts: En- suring scientific validity of feature-comparison methods. Washington, DC, USA: President’s Council of Advisors on Science and Technology (US); 2016

    2016

  18. [26]

    Reliability and Validity of Forensic Science Evidence

    Stern HS, Cuellar M, Kaye DH. Reliability and Validity of Forensic Science Evidence. Significance. 2019;16(2):21-4

    2019

  19. [27]

    Ongoing face recognition vendor test (FR VT) part 1: Verification

    Grother P, Ngan M, Hanaoka K. Ongoing face recognition vendor test (FR VT) part 1: Verification. National Institute of Standards and Technology. 2019

    2019

  20. [28]

    Accuracy and reliability of forensic latent fingerprint decisions

    Ulery BT, Hicklin RA, Buscaglia J, Roberts MA. Accuracy and reliability of forensic latent fingerprint decisions. Proceedings of the National Academy of Sciences. 2011;108(19):7733-8. Available from:https://www.pnas.org/doi/ abs/10.1073/pnas.1018707108. 23

    2011 doi

  21. [29]

    A study of false-positive and false-negative error rates in cartridge case comparisons

    Baldwin DP, Bajic SJ, Morris M, Zamzow D. A study of false-positive and false-negative error rates in cartridge case comparisons. Technical report; 2014

    2014

  22. [30]

    Methodological problems in every black-box study of forensic firearm comparisons

    Cuellar M, Vanderplas S, Luby A, Rosenblum M. Methodological problems in every black-box study of forensic firearm comparisons. Law, Probability and Risk. 2024 12;23(1):mgae015. Available from:https://doi.org/10.1093/lpr/ mgae015

    2024 doi

  23. [31]

    FRCSyn Challenge at W ACV 2024: Face Recognition Challenge in the Era of Synthetic Data

    Melzi P, Tolosana R, Vera-Rodriguez R, Kim M, Rathgeb C, Liu X, et al. FRCSyn Challenge at W ACV 2024: Face Recognition Challenge in the Era of Synthetic Data. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops; 2024. Available from:...

    2024 arXiv

  24. [32]

    The Impact of Balancing Real and Synthetic Data on Accuracy and Fairness in Face Recognition

    Atzori A, Cosseddu P, Fenu G, Marras M. The Impact of Balancing Real and Synthetic Data on Accuracy and Fairness in Face Recognition. arXiv preprint arXiv:240902867. 2024. Available from:https://arxiv.org/abs/2409.02867

    2024 arXiv

  25. [33]

    Synthetic to Authentic: Transferring Realism to 3D Face Renderings for Face Recognition

    Rahimi P, et al. Synthetic to Authentic: Transferring Realism to 3D Face Renderings for Face Recognition. In: Proceedings of the European Conference on Computer Vision (ECCV); 2024. Available from:https://arxiv.org/html/ 2407.07627v2

    2024 arXiv

  26. [34]

    Alias- free generative adversarial networks

    Karras T, Aittala M, Laine S, H¨ ark¨ onen E, Hellsten J, Lehtinen J, et al. Alias- free generative adversarial networks. Advances in neural information processing systems. 2021;34:852-63

    2021

  27. [35]

    Third time’s the charm? image and video editing with stylegan3

    Alaluf Y, Patashnik O, Wu Z, Zamir A, Shechtman E, Lischinski D, et al. Third time’s the charm? image and video editing with stylegan3. In: European Conference on Computer Vision. Springer; 2022. p. 204-20

    2022

  28. [36]

    Examining Pathological Bias in a Generative Adversarial Network Discriminator: A Case Study on a StyleGAN3 Model

    Grissom II A, Lei RF, Neto JFSR, Lin B, Trotter R. Examining Pathological Bias in a Generative Adversarial Network Discriminator: A Case Study on a StyleGAN3 Model. arXiv preprint arXiv:240209786. 2024

    2024

  29. [37]

    A style-based generator architecture for generative adversarial networks

    Karras T, Laine S, Aila T. A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition; 2019. p. 4401-10

    2019

  30. [38]

    Exposing GAN- generated profile photos from compact embeddings

    Mundra S, Porcile GJA, Marvaniya S, Verbus JR, Farid H. Exposing GAN- generated profile photos from compact embeddings. In: Proceedings of the 24 IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023. p. 884-92

    2023

  31. [39]

    Studying bias in gans through the lens of race

    Maluleke VH, Thakkar N, Brooks T, Weber E, Darrell T, Efros AA, et al. Studying bias in gans through the lens of race. In: European Conference on Computer Vision. Springer; 2022. p. 344-60

    2022

  32. [40]

    MRMCsamplesize: An R Package for Esti- mating Sample Sizes for Multi-Reader Multi-Case Studies

    Robert D, Sathyamurthy S, Putha P. MRMCsamplesize: An R Package for Esti- mating Sample Sizes for Multi-Reader Multi-Case Studies. medRxiv. 2023:2023- 09

    2023

  33. [41]

    Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation

    Karkkainen K, Joo J. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision; 2021. p. 1548-58

    2021

  34. [42]

    Gandiffface: Controllable generation of synthetic datasets for face recog- nition with realistic variations

    Melzi P, Rathgeb C, Tolosana R, Vera-Rodriguez R, Lawatsch D, Domin F, et al. Gandiffface: Controllable generation of synthetic datasets for face recog- nition with realistic variations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision; 2023. p. 3086-95

    2023

  35. [43]

    Age boundaries in juvenile justice systems

    Association NG, et al. Age boundaries in juvenile justice systems. Washington, DC: National Governors Association. 2021

    2021

  36. [44]

    Juvenile Offenders Ordinance, Cap

    Hong Kong SAR Government. Juvenile Offenders Ordinance, Cap. 226

  37. [45]

    Age of criminal responsibility; 2025

    UK Government. Age of criminal responsibility; 2025. Accessed: 2025-05-16. Available from:https://www.gov.uk/age-of-criminal-responsibility

    2025

  38. [46]

    Minimum age of criminal responsibility; 2025

    Attorney-General’s Department, Government of South Australia. Minimum age of criminal responsibility; 2025. Accessed: 2025-05-16. Avail- able from:https://www.agd.sa.gov.au/law-and-justice/consultation/ minimum-age-of-criminal-responsibility

    2025

  39. [47]

    2020 Census Redistricting Data (Public Law 94-171) Summary File; 2021

    U S Census Bureau. 2020 Census Redistricting Data (Public Law 94-171) Summary File; 2021. Accessed: 2025-05-07. Available from:https:// www.census.gov/data/datasets/2020/dec/2020-redistricting-data.html

    2020

  40. [48]

    The OpenCV Library

    Bradski G. The OpenCV Library. Dr Dobb’s Journal of Software Tools. 2000. 25

    2000

  41. [49]

    Blur-Generator: Simulates realistic blur on images using Python; 2023

    Lee N. Blur-Generator: Simulates realistic blur on images using Python; 2023. Accessed: 2025-05-07. Available from:https://github.com/NatLee/Blur- Generator

    2023

  42. [50]

    Interpreting the latent space of gans for semantic face editing

    Shen Y, Gu J, Tang X, Zhou B. Interpreting the latent space of gans for semantic face editing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition; 2020. p. 9243-52

    2020

  43. [51]

    LightFace: A Hybrid Deep Face Recognition Frame- work

    Serengil SI, Ozpinar A. LightFace: A Hybrid Deep Face Recognition Frame- work. In: 2020 Innovations in Intelligent Systems and Applications Confer- ence (ASYU). IEEE; 2020. p. 23-7. Available from:https://doi.org/10.1109/ ASYU50717.2020.9259802

    2020

  44. [52]

    Retinaface: Single-shot multi- level face localisation in the wild

    Deng J, Guo J, Ververas E, Kotsia I, Zafeiriou S. Retinaface: Single-shot multi- level face localisation in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition; 2020. p. 5203-12

    2020

  45. [53]

    Arcface: Additive angular margin loss for deep face recognition

    Deng J, Guo J, Xue N, Zafeiriou S. Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition; 2019. p. 4690-9

    2019

  46. [54]

    Face recognition oak ridge (faro): A framework for distributed and scalable biometrics applications

    Bolme DS, Srinivas N, Brogan J, Cornett D. Face recognition oak ridge (faro): A framework for distributed and scalable biometrics applications. In: 2020 IEEE International Joint Conference on Biometrics (IJCB). IEEE; 2020. p. 1-8

    2020

  47. [55]

    FR VT 1:1 Verification Visa Results; n.d

    of Standards NI, Technology. FR VT 1:1 Verification Visa Results; n.d. Accessed: 2025-05-16.https://pages.nist.gov/frvt/plots/11/visa.html

    2025

  48. [56]

    NA, editor

    Grother P, Ngan M, Hanaoka K. NA, editor. Face Recognition Vendor Test (FR VT), Part 3: Demographic Effects. NA; 2019

    2019

  49. [57]

    Replicating Gimp’s brightness/contrast adjustment in python

    Greenawalt J. Replicating Gimp’s brightness/contrast adjustment in python. Medium. 2023 21 December. Available from:https: //medium.com/@jsgreenawalt/replicating-gimps-brightness-contrast- adjustment-in-python-e9c962391162. 26

    2023

  50. [2022]

    Point in time: 01/07/2022.https://www.elegislation.gov.hk/hk/ cap226?xpid=ID 1438402856544 001

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.