Pith. sign in

REVIEW 4 major objections 5 minor 49 references

TrustSkin: A Fairness Pipeline for Trustworthy Facial Affect Analysis Across Skin Tone

T0 review · 4 major / 5 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 The skin-tone scale chosen for an emotion-AI audit can hide bias against darker-skinned faces.

desk verdict A useful first comparison of skin-tone taxonomies in FAA, but the central F1-disparity claim reverses when recomputed from the paper's own Table IV. read the letter →

arxiv 2505.20637 v1 pith:UBAUTR4W submitted 2025-05-27 cs.CV cs.AI

Ana M. Cabanas , Alma Pedro , Domingo Mery This is my paper · ORCID
classification cs.CVcs.AI
keywords facialaffectanalysisskintonefairnessIndividualTypologyAnglehue-lightnessclassificationNetEqualOpportunitybiasmeasurementmodelinterpretability
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

The paper claims that fairness audits of facial affect analysis are sensitive to how skin tone is measured, and that the standard Individual Typology Angle (ITA) can hide disparities affecting darker-skinned people. Comparing ITA with a perceptually grounded lightness-and-hue ($H^*$-$L^*$) classification on AffectNet with a MobileNet emotion classifier—where dark faces are only about 2% of the test set—the paper finds that the two taxonomies group faces differently and yield different fairness numbers: $H^*$-$L^*$ shows larger F1-score and true-positive-rate gaps across skin-tone groups, plus different Equal Opportunity patterns. The paper interprets the larger gaps as better measurement rather than worse model behavior, because ITA misassigns some medium and light faces to the dark group under variable lighting. If correct, fairness studies that rely only on ITA may underestimate how much darker-skinned users are disadvantaged by emotion recognition systems.

What carries the argument

The load-bearing machinery is a pair of skin-tone classifiers and the fairness metrics that compare them. ITA, defined as $\mathrm{ITA} = \arctan((L^*-50)/b^*) \cdot 180/\pi$, uses lightness and blue-yellow chroma but drops the green-red axis ($a^*$). The alternative $H^*$-$L^*$ method uses lightness $L^*$ plus hue angle $\mathrm{Hue} = \arctan(b^*/a^*)$, with thresholds (Light $L^*>67$, Medium $37 \leq L^* \leq 67$, Dark $L^*<37$) and a brown-tone override rule that reclassifies pixels with certain RGB ranges and hue between $20^\circ$ and $50^\circ$ as Dark regardless of $L^*$. That override is what lets the method catch dark-skinned faces whose lightness was raised by lighting, and the fairness metrics—F1 gap, accuracy range, and Equal Opportunity difference—turn each taxonomy into quantitative audit conclusions.

What would settle it

Collect a sample of faces with measured skin tone (e.g., spectrophotometer or self-report), compare ITA and hue-lightness labels against those measurements, and recompute the fairness gaps using only faces both methods label correctly; if ITA agrees with measured tone as often as hue-lightness on dark faces, the larger hue-lightness disparity would be a grouping artifact rather than evidence of hidden bias.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the skin-tone taxonomy used for fairness evaluation is itself a source of variation in the conclusions. On the AffectNet test set, the $H^*$-$L^*$ grouping produces an F1-score disparity of $0.080$ and a true-positive-rate disparity of $0.106$, while ITA grouping produces $0.050$ and $0.091$, respectively, and the Equal Opportunity heatmaps differ qualitatively. The paper attributes the difference to ITA pulling medium and light faces into the dark group under illumination changes, so that ITA's dark-group recall is partly inflated; by including hue and a brown-tone override, $H^*$-$L^*$ isolates true dark faces and thereby exposes larger, previously masked disparities. The paper concludes that $H^*$-$L^*$ gives more consistent subgrouping and that ITA-based evaluations may overlook disparities affecting darker-skinned individuals.

Load-bearing premise

The whole conclusion depends on the assumption that the hue-lightness method labels dark skin more accurately than ITA does, because the rule that overrides lightness for brown tones was tuned on two borderline cases from the same dataset rather than checked against real ground-truth skin tone.

Editorial extensions

If this is right

  • Fairness audits of facial affect systems that use only ITA should be rechecked with a hue-aware measure, because the choice of taxonomy changes which groups look disadvantaged.
  • The severe underrepresentation of dark skin in AffectNet (roughly 2% of samples) means reported dark-group metrics are statistically fragile and should be treated as diagnostic hints rather than stable estimates.
  • The larger F1 and TPR disparities seen with $H^*$-$L^*$ should be read as increased visibility of bias, not as evidence that the model treats dark skin worse than ITA suggested.
  • Equal Opportunity diagnostics can help separate real subgroup gaps from grouping artifacts, since ITA's dark-group recall is partly inflated by misclassified medium and light faces.
  • Rule-based corrections like the brown-tone override improve subgroup fidelity but need dataset-specific tuning before being reused.

Reading between the lines

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

  • Editorial inference: if hue-aware measures become standard, previously published ITA-based fairness conclusions in emotion recognition may need to be revisited, and some audits that looked acceptable could flip to showing bias.
  • Editorial inference: the same measurement-choice effect likely extends beyond emotion recognition to face recognition and dermatology datasets that rely on colorimetric skin-tone proxies, so the masking may be widespread.
  • Editorial inference: a direct test of the paper's interpretation would be to collect spectrophotometer or self-reported skin-tone labels for a sample of faces and compare ITA and $H^*$-$L^*$ agreement; if ITA labels are equally accurate for dark skin, the larger $H^*$-$L^*$ disparity would be a grouping artifact rather than a revelation.
  • Editorial inference: re-running the audit on a balanced test set with many more dark-skin samples would show whether the larger $H^*$-$L^*$ disparity persists or shrinks into a small-sample artifact.
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 / 5 minor

Summary. This paper compares two objective skin-tone stratification methods—ITA and an L*–H* (Hue–Lightness) method—for fairness evaluation of facial affect analysis. Using AffectNet and a MobileNet classifier, it reports underrepresentation of dark skin tones, per-group emotion metrics, F1/TPR disparities, Equal Opportunity analysis, and Grad-CAM visualizations. The central claim is that the choice of skin-tone measurement changes fairness conclusions and that ITA-based audits may underestimate disparities affecting darker-skinned individuals.

Significance. The paper addresses a timely and under-explored question: whether the operationalization of skin tone changes fairness audits in facial affect analysis. Its proposed modular pipeline, explicit comparison of ITA and H*L*, and use of Equal Opportunity and Grad-CAM are useful ingredients. The authors are candid about limitations such as empirical thresholds and the tiny dark-skin sample. However, the quantitative support for the headline claim is undermined by an internal inconsistency between Figure 6 and Table IV, and by the lack of external validation for the labeling rule. If the numerical issues are resolved, the paper could make a modest but useful contribution; in its current form the central quantitative claim is not reproducible from the supplied data.

major comments (4)
  1. [Section IV, Figure 6 vs. Table IV] The F1-disparity values in Figure 6 do not reproduce from the paper's own Table IV under the definition in Section II.C. Averaging the eight per-emotion F1 scores in Table IV gives H*L* Light=0.414, Medium=0.388, Dark=0.398 (gap 0.026), while ITA Light=0.403, Medium=0.394, Dark=0.478 (gap 0.084). Thus Table IV implies the opposite ordering from Figure 6: ITA has the larger macro-F1 gap, and the Dark group is the best ITA group, not a disadvantaged one. This directly affects the abstract's 'up to 0.08' claim and the Section IV interpretation that H*L* has a more pronounced F1 disparity. Please recompute and correct Figure 6 and the accompanying text, or report the exact computation that produced the printed values.
  2. [Section II.B] The brown-tone override is introduced with thresholds for RGB ranges and hue, and the only evidence offered is that it 'correctly reassigning two borderline cases' in the same dataset. Since the same override is then used to argue that H*L* gives more accurate dark-skin labels and hence better fairness diagnostics, this is in-sample validation. An independent ground truth (spectrophotometer, self-report, or manually labeled held-out set) or at least a sensitivity analysis over the override thresholds is needed before the claim that ITA underestimates dark-skin disparities can be supported.
  3. [Section III.A, Tables II-IV] The Dark group contains only 52 test images, yet the paper reports per-emotion precision/recall/F1 without confidence intervals. Values such as Dark H*L* Disgust precision 1.00, Surprise precision 0.12, and Neutral recall 1.00 are likely based on very small denominators and should be accompanied by counts, confidence intervals, or bootstrap estimates. Without this, the strong qualitative claims about Dark-group performance in Section III.A and Section IV.A are not statistically grounded.
  4. [Abstract / Section I vs. Section V] The abstract and Section I state that 'targeted data augmentation failed to improve fairness,' but Section V and the Ethical Impact Statement say that no dataset rebalancing was applied in this work. These statements cannot both describe the executed experiments. Please clarify whether augmentation experiments were run and, if so, report them; otherwise remove the augmentation claim from the abstract and introduction.
minor comments (5)
  1. [Tables II-IV] The emotion row uses 'Neutre' instead of 'Neutral'; fix the typo for consistency with the rest of the paper.
  2. [Figure 6] The caption appears incomplete ('ITA ITAH*-L* H*-L*') and the panels are unlabeled; add a clear caption and axis labels so the reader can distinguish the F1 and TPR panels.
  3. [Section II.B] The Hue formula is written as arctan(b*/a*); specify that a four-quadrant arctangent (atan2) is used, since a* can be negative and the angle range matters for the stated 20–50 degree interval.
  4. [Throughout] The notation 'H*L*' and 'L*H*' is used interchangeably; choose one form and use it consistently.
  5. [Section II.A] The paper refers to 'validated thresholds from prior work [16], [31]' but does not list the actual segmentation thresholds; a short appendix with the YCrCb and HSV thresholds would improve reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

Brown-tone override is fit to two same-dataset cases and then cited as evidence that H*L* is more reliable; the fairness-gap comparison itself is not circularly fit.

  1. fitted input called prediction [Section II.B (brown-tone override) and Section IV.A (interpretation of Dark-group recall)]
    "To address this, we apply a brown-tone override that reclassifies pixels as Dark based on RGB ranges and Hue, independently of their initial L∗ label. Specifically, a pixel qualifies if: 100 ≤ R ≤ 170, 60 ≤ G ≤ 110, 40 ≤ B ≤ 85, with R > G > B, and (R − G) < 30, (G − B) < 25. The Hue angle must fall within 20◦ ≤ H ∗ ≤ 50◦, capturing typical brown hues. This rule improved classification in underrepresented groups, correctly reassigning two borderline cases (see last row of Figure 4)."

    The RGB and hue thresholds of the brown-tone override were explicitly chosen so that two AffectNet borderline cases would be reassigned to the Dark group. The paper then presents this same reassignment as evidence that the H*L* method 'improves tone estimation' and 'more reliably captures chromatic characteristics, particularly brown tones.' The evidence for H*L* superiority on dark skin is therefore the in-sample fit used to define the method: the rule is constructed to classify those cases as Dark, and the same cases are then displayed as proof of improved classification.

full rationale

The paper's central contribution is an empirical comparison of fairness metrics under two skin-tone taxonomies. That comparison is not circular: the F1 and TPR disparities are computed from model predictions and the two taxonomies, and the paper does not fit the taxonomy to produce a desired fairness gap. The clearest circular element is the brown-tone override, whose thresholds were tuned to two borderline cases in the same dataset and then cited as evidence that the H*L* method is more accurate on dark skin. The authors acknowledge this limitation, calling the thresholds 'empirical' and 'dataset-specific,' which makes the practice transparent but does not remove the in-sample nature of that particular claim. No load-bearing self-citation was found; the Mery citations are only offered as future explainability alternatives. The reported inconsistency between Figure 6's F1 disparity values and Table IV's per-group F1 scores, and the abstract's claim about augmentation versus the statement that no rebalancing was applied, are reproducibility and internal-consistency concerns rather than circularity, and I do not count them as circular steps.

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

The paper's central comparison depends on a set of hand-chosen thresholds and an override rule that is tuned in-sample. There are no new postulated entities, but several domain assumptions about skin segmentation and the reliability of Lab color features are load-bearing.

free parameters (3)
  • ITA thresholds = Light >55 degrees, Medium 30-55 degrees, Dark <30 degrees
    Chosen by the authors to collapse six Fitzpatrick levels into three groups; thresholds are from prior conventions but are discretionary.
  • H*L* L* thresholds = Light >67.0, Medium 37.0-67.0, Dark <37.0
    Chosen to define skin tone groups in the hue-lightness method; no derivation from data.
  • Brown-tone override thresholds = R 100-170, G 60-110, B 40-85, R>G>B, (R-G)<30, (G-B)<25, Hue 20-50 degrees
    Hand-tuned to reclassify two borderline dark-skin examples in the dataset; authors note it may need tuning for other datasets.
assumptions (4)
  • domain assumption The skin pixel segmentation (Otsu plus chrominance filters in YCrCb and HSV) correctly isolates facial skin regions.
    Assumed from prior work; no validation on AffectNet or reporting of failure cases.
  • domain assumption Average RGB of segmented skin pixels, converted to Lab, is a reliable proxy for skin tone.
    Ignores lighting, sensor, and shading variations; central to both classification methods.
  • domain assumption The AffectNet test set is an appropriate real-world distribution for fairness evaluation.
    Used to justify computing fairness on the unbalanced test set, but the 52-image dark group limits statistical power.
  • domain assumption Grad-CAM attention differences by group indicate variation in feature encoding relevant to fairness.
    Post-hoc saliency maps are coarse and may not reflect true decision logic; the paper itself concedes this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TrustSkin: A Fairness Pipeline for Trustworthy Facial Affect Analysis Across Skin Tone." pith.science (2026). https://pith.science/paper/UBAUTR4W

@misc{pith2026250520637,
  author       = {Pith},
  title        = {Pith review of: TrustSkin: A Fairness Pipeline for Trustworthy Facial Affect Analysis Across Skin Tone},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBAUTR4W}},
  note         = {Machine review of arXiv:2505.20637}
}
abstract

Understanding how facial affect analysis (FAA) systems perform across different demographic groups requires reliable measurement of sensitive attributes such as ancestry, often approximated by skin tone, which itself is highly influenced by lighting conditions. This study compares two objective skin tone classification methods: the widely used Individual Typology Angle (ITA) and a perceptually grounded alternative based on Lightness ($L^*$) and Hue ($H^*$). Using AffectNet and a MobileNet-based model, we assess fairness across skin tone groups defined by each method. Results reveal a severe underrepresentation of dark skin tones ($\sim 2 \%$), alongside fairness disparities in F1-score (up to 0.08) and TPR (up to 0.11) across groups. While ITA shows limitations due to its sensitivity to lighting, the $H^*$-$L^*$ method yields more consistent subgrouping and enables clearer diagnostics through metrics such as Equal Opportunity. Grad-CAM analysis further highlights differences in model attention patterns by skin tone, suggesting variation in feature encoding. To support future mitigation efforts, we also propose a modular fairness-aware pipeline that integrates perceptual skin tone estimation, model interpretability, and fairness evaluation. These findings emphasize the relevance of skin tone measurement choices in fairness assessment and suggest that ITA-based evaluations may overlook disparities affecting darker-skinned individuals.

Figures

Figures reproduced from arXiv: 2505.20637 by the authors.

Figure 1
Figure 1. Distribution of ITA values in the cleaned AffectNet training set. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Skin tone distribution in the (H∗–L∗) color space. Point colors correspond to their respective RGB values. Dashed horizontal lines denote L∗ thresholds for skin tone classification: Light (L∗ > 67.0), Medium (37.0 ≤ L∗ ≤ 67.0), and Dark (L∗ < 37.0) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Proposed classification using L∗ and Hue (H∗ = arctan(b ∗/a∗)), capturing both lightness and chromaticity. Images were randomly sampled from the training subset of AffectNet. This method improves tone estima￾tion, particularly under variable lighting and darker skin tones. C. Fairness Evaluation Framework To evaluate how facial affect recognition performance varies across skin tone groups, we adopt a multifaceted fa… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Distribution of labeled emotions in the AffectNet test set by skin [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Comparison of F1-score and TPR disparity (fairness gap) between [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Grad-CAM visualizations of facial emotion predictions across skin [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Overview of the proposed fairness-aware FAA pipeline. The process begins with a facial dataset that undergoes preprocessing and objective [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 48 canonical work pages

  1. [1]

    Demographic Bias in Biometrics: A Survey on an Emerging Chal- lenge,

    P. Drozdowski, C. Rathgeb, A. Dantcheva, N. Damer, and C. Busch, “Demographic Bias in Biometrics: A Survey on an Emerging Chal- lenge,” IEEE Transactions on Technology and Society , vol. 1, no. 2, pp. 89–103, 2020

    work page 2020

  2. [2]

    The Impact of Racial Distribution in Training Data on Face Recognition Bias: A Closer Look,

    M. Kolla and A. Savadamuthu, “The Impact of Racial Distribution in Training Data on Face Recognition Bias: A Closer Look,” in IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, WACVW 2023, pp. 313–322, 2023

    work page 2023

  3. [3]

    NIST special publication 1270. Towards a standard for identifying and managing bias in AI,

    R. Schwartz, A. Vassilev, K. Greene, L. Perine, A. Burt, P. Hall, and K. Greene, “NIST special publication 1270. Towards a standard for identifying and managing bias in AI,” NIST special, pp. 1–86, 2022

    work page 2022

  4. [4]

    Not in My Face: Challenges and Ethical Considerations in Automatic Face Emotion Recognition Technology,

    M. Mattioli and F. Cabitza, “Not in My Face: Challenges and Ethical Considerations in Automatic Face Emotion Recognition Technology,” Machine Learning and Knowledge Extraction, vol. 6, no. 4, pp. 2201– 2231, 2024

    work page 2024

  5. [5]

    Investigating Bias and Fairness in Facial Expression Recognition,

    T. Xu, J. White, S. Kalkan, and H. Gunes, “Investigating Bias and Fairness in Facial Expression Recognition,” in Computer Vision – ECCV 2020 Workshops (A. Bartoli and A. Fusiello, eds.), (Cham), pp. 506–523, Springer International Publishing, 2020

    work page 2020

  6. [6]

    Balancing the Scales: Enhanc- ing Fairness in Facial Emotion Recognition with Latent Alignment,

    S. S. A. Rizvi, A. Seth, and P. Narang, “Balancing the Scales: Enhanc- ing Fairness in Facial Emotion Recognition with Latent Alignment,” in Pattern Recognition (A. Antonacopoulos, S. Chaudhuri, R. Chellappa, C.-L. Liu, S. Bhattacharya, and U. Pal, eds.), (Cham), pp. 113–128, Springer Nature Switzerland, 2025

    work page 2025

  7. [7]

    Faces of Fairness: Examining Bias in Facial Expression Recognition Datasets and Mod- els,

    M. M. Hosseini, A. P. Fard, and M. H. Mahoor, “Faces of Fairness: Examining Bias in Facial Expression Recognition Datasets and Mod- els,” pp. 1–14, 2025

    work page 2025

  8. [8]

    Notes from the AI frontier : Tackling bias in AI ( and in humans ),

    J. Silberg and J. Manyika, “Notes from the AI frontier : Tackling bias in AI ( and in humans ),” Mckinsey Global Institute , pp. 1–8, 2019

    work page 2019

Show all 49 references
  1. [9]

    Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,

    J. Buolamwini and T. Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” Proceedings of Machine Learning Research , vol. 81, pp. 77–91, 2018

    2018

  2. [10]

    How We’ve Taught Algorithms to See Identity: Constructing Race and Gender in Image Databases for Facial Analysis,

    M. K. Scheuerman, K. Wade, C. Lustig, and J. R. Brubaker, “How We’ve Taught Algorithms to See Identity: Constructing Race and Gender in Image Databases for Facial Analysis,” Proceedings of the ACM on Human-Computer Interaction , vol. 4, no. CSCW1, pp. 1–35, 2020

    2020

  3. [11]

    One label, one billion faces: Usage and consistency of racial categories in computer vision,

    Z. Khan and Y . Fu, “One label, one billion faces: Usage and consistency of racial categories in computer vision,” FAccT 2021 - Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 587–597, 2021

    2021

  4. [12]

    Consensus and Subjectivity of Skin Tone Annotation for ML Fairness,

    C. Schumann, G. O. Olanubi, A. Wright, E. Monk, C. Heldreth, and S. Ricco, “Consensus and Subjectivity of Skin Tone Annotation for ML Fairness,” Advances in Neural Information Processing Systems , vol. 36, no. NeurIPS, 2023

    2023

  5. [13]

    Which Skin Tone Measures Are the Most Inclusive? An Investigation of Skin Tone Measures for Artificial Intelligence,

    C. M. Heldreth, E. P. Monk, A. T. Clark, C. Schumann, X. Eyee, and S. Ricco, “Which Skin Tone Measures Are the Most Inclusive? An Investigation of Skin Tone Measures for Artificial Intelligence,” ACM Journal on Responsible Computing , vol. 1, no. 1, pp. 1–21, 2024

    2024

  6. [14]

    The validity and practicality of sun-reactive skin types I through VI.,

    T. B. Fitzpatrick, “The validity and practicality of sun-reactive skin types I through VI.,” Archives of dermatology, vol. 124, pp. 869–871, jun 1988

    1988

  7. [15]

    Fitz- patrick Skin Type, Individual Typology Angle, and Melanin Index in an African Population: Steps Toward Universally Applicable Skin Photosensitivity Assessments.,

    M. Wilkes, C. Y . Wright, J. L. du Plessis, and A. Reeder, “Fitz- patrick Skin Type, Individual Typology Angle, and Melanin Index in an African Population: Steps Toward Universally Applicable Skin Photosensitivity Assessments.,” JAMA dermatology, vol. 151, pp. 902– 903, aug 2015

    2015

  8. [16]

    Diversity in Faces,

    M. Merler, N. Ratha, R. S. Feris, and J. R. Smith, “Diversity in Faces,” pp. 1–29, 2019

    2019

  9. [17]

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

    K. Karkkainen and J. Joo, “FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation,” Proceedings - 2021 IEEE Winter Conference on Applications of Computer Vision, WACV 2021 , pp. 1547–1557, 2021

    2021

  10. [18]

    Beyond Skin Tone: A Multidi- mensional Measure of Apparent Skin Color,

    W. Thong, P. Joniak, and A. Xiang, “Beyond Skin Tone: A Multidi- mensional Measure of Apparent Skin Color,” Proceedings of the IEEE International Conference on Computer Vision , pp. 4880–4890, 2023

    2023

  11. [19]

    Racial Bias within Face Recognition : A Survey,

    T. P. Breckon, “Racial Bias within Face Recognition : A Survey,” vol. 1, no. 1, 2023

    2023

  12. [20]

    Evaluating Group Fairness in News Recommendations : A Comparative Study of Algorithms and Metrics,

    B. Huebner, T. E. Kolb, and J. Neidhardt, “Evaluating Group Fairness in News Recommendations : A Comparative Study of Algorithms and Metrics,” pp. 337–346

  13. [21]

    Meta Balanced Network for Fair Face Recognition,

    M. Wang, Y . Zhang, and W. Deng, “Meta Balanced Network for Fair Face Recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, pp. 8433–8448, 2021

    2021

  14. [22]

    Towards Measuring Fairness in AI: The Casual Conversations Dataset,

    C. Hazirbas, J. Bitton, B. Dolhansky, J. Pan, A. Gordo, and C. C. Ferrer, “Towards Measuring Fairness in AI: The Casual Conversations Dataset,” IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 4, no. 3, pp. 324–332, 2022

    2022

  15. [23]

    Racial Influence on Automated Perceptions of Emotions,

    L. Rhue, “Racial Influence on Automated Perceptions of Emotions,” SSRN Electronic Journal , 2018

    2018

  16. [24]

    Addressing Racial Bias in Facial Emotion Recognition,

    A. Fan, X. Xiao, and P. Washington, “Addressing Racial Bias in Facial Emotion Recognition,” 2023

    2023

  17. [25]

    Understanding perception of algorithmic decisions: Fair- ness, trust, and emotion in response to algorithmic management,

    M. K. Lee, “Understanding perception of algorithmic decisions: Fair- ness, trust, and emotion in response to algorithmic management,” Big data & society , vol. 5, no. 1, p. 2053951718756684, 2018

    2018

  18. [26]

    Counterfactual fairness for facial expression recognition,

    J. Cheong, S. Kalkan, and H. Gunes, “Counterfactual fairness for facial expression recognition,” in European Conference on Computer Vision, pp. 245–261, Springer, 2022

    2022

  19. [27]

    Bias and fairness on multimodal emotion detection algorithms,

    M. Schmitz, R. Ahmed, and J. Cao, “Bias and fairness on multimodal emotion detection algorithms,” arXiv preprint arXiv:2205.08383 , 2022

    2022 arXiv

  20. [28]

    Domain adaptation for bias mitigation in affective computing: use cases for facial emotion recognition and sentiment analysis systems,

    P. Singhal, S. Gokhale, A. Shah, D. K. Jain, R. Walambe, A. Ekart, and K. Kotecha, “Domain adaptation for bias mitigation in affective computing: use cases for facial emotion recognition and sentiment analysis systems,” Discover Applied Sciences , vol. 7, no. 4, p. 229, 2025

    2025

  21. [29]

    Be- yond Accuracy: Fairness, Scalability, and Uncertainty Considerations in Facial Emotion Recognition,

    L. Fromberg, T. Nielsen, F. D. Frumosu, and L. H. Clemmensen, “Be- yond Accuracy: Fairness, Scalability, and Uncertainty Considerations in Facial Emotion Recognition,” in Northern Lights Deep Learning Conference, pp. 67–74, PMLR, 2024

    2024

  22. [30]

    Com- putational methods for pigmented skin lesion classification in images: review and future trends,

    R. B. Oliveira, J. P. Papa, A. S. Pereira, and J. M. R. Tavares, “Com- putational methods for pigmented skin lesion classification in images: review and future trends,”Neural Computing and Applications, vol. 29, no. 3, pp. 613–636, 2018

    2018

  23. [31]

    Human Skin Detection Using RGB, HSV and YCbCr Color Models,

    S. Kolkur, D. Kalbande, P. Shimpi, C. Bapat, and J. Jatakia, “Human Skin Detection Using RGB, HSV and YCbCr Color Models,” vol. 137, pp. 324–332, 2017

    2017

  24. [32]

    Colorimetric skin tone scale for improved accuracy of human skin tone annotations.,

    C. Cook, J. Howard, L. Rabbitt, I. Shuggi, Y . Sirotin, J. Tipton, and A. Vemury, “Colorimetric skin tone scale for improved accuracy of human skin tone annotations.,” ACM J. Responsib. Comput. , May

  25. [33]

    Fairness and Bias Mitigation in Computer Vision: A Survey,

    S. Dehdashtian, R. He, Y . Li, G. Balakrishnan, N. Vasconcelos, V . Ordonez, and V . N. Boddeti, “Fairness and Bias Mitigation in Computer Vision: A Survey,” 2024

    2024

  26. [34]

    AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild,

    A. Mollahosseini, B. Hasani, and M. H. Mahoor, “AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild,” IEEE Transactions on Affective Computing , vol. 10, no. 1, pp. 18–31, 2019

    2019

  27. [35]

    Exploring strategies to generate Fitz- patrick skin type metadata for dermoscopic images using individual ty- pology angle techniques,

    A. Corbin and O. Marques, “Exploring strategies to generate Fitz- patrick skin type metadata for dermoscopic images using individual ty- pology angle techniques,” Multimedia Tools and Applications, vol. 82, no. 15, pp. 23771–23795, 2023

    2023

  28. [37]

    Variations in skin colour and the biological consequences of ultraviolet radiation exposure,

    S. Del Bino and F. Bernerd, “Variations in skin colour and the biological consequences of ultraviolet radiation exposure,” British Journal of Dermatology , vol. 169, no. SUPPL. 3, pp. 33–40, 2013

    2013

  29. [38]

    Achieve fairness without demographics for dermatological disease diagnosis,

    C. H. Chiu, Y . J. Chen, Y . Wu, Y . Shi, and T. Y . Ho, “Achieve fairness without demographics for dermatological disease diagnosis,” Medical Image Analysis, vol. 95, no. December 2023, p. 103188, 2024

    2023

  30. [39]

    Estimating Skin Tone and Effects on Classification Performance in Dermatology Datasets,

    N. M. Kinyanjui, T. Odonga, C. Cintas, N. C. F. Codella, R. Panda, P. Sattigeri, and K. R. Varshney, “Estimating Skin Tone and Effects on Classification Performance in Dermatology Datasets,” pp. 1–11, 2019

    2019

  31. [40]

    Towards Fairness in AI for Melanoma Detection: Systemic Review and Recommenda- tions,

    L. N. Montoya, J. S. Roberts, and B. S. Hidalgo, “Towards Fairness in AI for Melanoma Detection: Systemic Review and Recommenda- tions,” 2024

    2024

  32. [41]

    AI Fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias,

    R. K. Bellamy, A. Mojsilovic, S. Nagar, K. N. Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, Y . Zhang, K. Dey, M. Hind, S. C. Hoffman, S. Houde, K. Kannan, P. Lohia, J. Martino, and S. Mehta, “AI Fairness 360: An extensible toolkit for detecting and...

    2019

  33. [42]

    FACET: Fairness in Computer Vision Evalua- tion Benchmark,

    L. Gustafson, C. Rolland, N. Ravi, Q. Duval, A. Adcock, C. Y . Fu, M. Hall, and C. Ross, “FACET: Fairness in Computer Vision Evalua- tion Benchmark,” Proceedings of the IEEE International Conference on Computer Vision , pp. 20313–20325, 2023

    2023

  34. [43]

    Equality of opportunity in super- vised learning,

    M. Hardt, E. Price, and N. Srebro, “Equality of opportunity in super- vised learning,” Advances in Neural Information Processing Systems , pp. 3323–3331, 2016

    2016

  35. [44]

    Coroama and A

    L. Coroama and A. Groza, Evaluation Metrics in Explainable Artificial Intelligence (XAI), vol. 1675 CCIS. Springer International Publishing, 2022

    2022

  36. [45]

    Explaining Face Recognition Through SHAP-Based Pixel-Level Face Image Quality Assessment,

    C. Biagi, L. Rethfeld, A. Kuijper, and P. Terh ¨orst, “Explaining Face Recognition Through SHAP-Based Pixel-Level Face Image Quality Assessment,” in 2023 IEEE International Joint Conference on Bio- metrics (IJCB), pp. 1–10, 2023

    2023

  37. [46]

    On Black-Box Explanation for Face Ver- ification,

    D. Mery and B. Morris, “On Black-Box Explanation for Face Ver- ification,” 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2022 , pp. 1194–1203, 2022

    2022

  38. [47]

    True Black-Box Explanation in Facial Analysis,

    D. Mery, “True Black-Box Explanation in Facial Analysis,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 1595–1604, 2022

    2022

  39. [48]

    Grad-cam: Why did you say that? visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Why did you say that? visual explanations from deep networks via gradient-based localization,” Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization , vol. 17, pp....

    2016

  40. [49]

    On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation,

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. M ¨uller, and W. Samek, “On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation,”PLOS ONE, vol. 10, no. 7, pp. 1–46, 2015

    2015

  41. [50]

    Focused LRP: Explainable AI for Face Morphing Attack Detection,

    C. Seibold, A. Hilsmann, and P. Eisert, “Focused LRP: Explainable AI for Face Morphing Attack Detection,” Proceedings - 2021 IEEE Winter Conference on Applications of Computer Vision Workshops, WACVW 2021, pp. 88–96, 2021

    2021

Pith tools

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