Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Balancing Beyond Discrete Categories: Continuous Demographic Labels for Fair Face Recognition

T0 review · 4 major / 6 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 Treating ethnicity as a continuous score per identity, and balancing face datasets on that score, yields fairer models than balancing discrete group counts.

desk verdict A genuinely useful empirical study of continuous-ethnicity balancing, but the in-sample pseudo-label makes the core comparison over random sampling look like data cleaning. read the letter →

arxiv 2506.01532 v4 pith:JRH2A4ED submitted 2025-06-02 cs.CV

Pedro C. Neto , Naser Damer , Jaime S. Cardoso , Ana F. Sequeira This is my paper · ORCID
classification cs.CV
keywords facerecognitiondemographicbiasfairnesscontinuousethnicitylabelsdatasetbalancingidentityrepresentativenesssamplingprotocolsmitigation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that ethnicity in face recognition should be treated as a continuous spectrum rather than a discrete category. The authors build a classifier on top of a face recognition backbone to score how strongly each image, and each identity, represents its labeled ethnic group, and they show that identities within the same category vary widely on this score. They propose three sampling protocols that remove identities so that the remaining data is balanced in this continuous space, and they train over sixty-five models to compare those subsets with randomly downsampled subsets that keep discrete counts equal. Their central claim is that continuous-space balancing consistently outperforms discrete balancing: lower cross-ethnicity accuracy gaps (STD and SER) with equal or better average accuracy, even when up to half of all identities are removed. If correct, this reframes dataset balance, because equal counts per ethnicity no longer count as balanced when individuals differ in representativeness.

What carries the argument

The carrying mechanism is a four-dimensional continuous ethnicity score produced by a linear classifier layer trained on top of a frozen face recognition backbone, which assigns each image a probability vector over four ethnic groups. Three aggregation levels link images to identities to ethnicities: the image score $ies_{j,i}$, the identity score $IDS_j$ (in Protocol A the mean of its image scores, in Protocols B and C the sum), and the ethnicity score $ES_y$ (mean over identities in A and B, sum in C). Three removal protocols repeatedly delete the least representative identity from the ethnicity with the lowest score (A and B) or the highest score (C), creating a sequence of continuously balanced subsets; a relabeling variant reassigns each identity to its maximum-scoring ethnicity to show the effect is not merely the removal of mislabeled samples.

What would settle it

Train the same three sampling protocols with a continuous ethnicity score obtained from human-annotated ancestry proportions instead of classifier confidence, and compare against randomly downsampled discrete-balanced subsets on the per-ethnicity benchmark. If the fairness and accuracy advantage over random sampling disappears, then the reported gains came from removing hard or in-sample images rather than from the continuous nature of ethnicity; a cheaper confound check is to inspect whether identities removed by Protocol A are systematically low-quality according to a face quality score.

Watch

Extended reading notes

Core claim

The central discovery is that the discrete ethnicity label attached to an identity does not measure that identity's contribution to demographic balance. The paper defines a continuous ethnicity score at three levels: per-image score $ies$, per-identity score $IDS$, and per-ethnicity score $ES$, computed by a classifier attached to a pretrained face recognition model. It then establishes three results: removing the highest-scoring identities from one ethnicity degrades accuracy on that ethnicity much faster than removing the lowest-scoring ones; a dataset with equal identity counts per ethnicity is still visibly imbalanced in the continuous space; and models trained on subsets balanced by continuous scores sit on or near the Pareto frontier of error versus fairness, dominating randomly sampled discrete-balanced subsets of the same size. The authors take this as evidence that balancing should equalize representativeness scores rather than category counts, and that this reframing holds across several network sizes and two training losses.

Load-bearing premise

The central assumption is that the classifier's confidence score actually measures how representative an identity is of its ethnic group; the paper itself notes the score only captures how easily an image is classified into that group, so if the score is really measuring image difficulty or training-set memorization, the fairness gains could be a data-cleaning artifact rather than evidence about continuous ethnicity.

Editorial extensions

If this is right

  • A training set with an equal number of identities per ethnicity can still be demographically imbalanced, so future dataset construction should report continuous-score distributions, not just counts.
  • Under Protocol B, removing 25% of identities costs almost no accuracy while clearly improving fairness, meaning fairer models can be trained on smaller, cheaper datasets.
  • Protocol A produces the largest fairness gains at extreme removal levels, cutting STD from 1.62 to 0.55 when half of the identities are removed while keeping average accuracy at the random-sampling level.
  • The advantage of continuous balancing persists across iResNet-34, iResNet-50, and iResNet-100 and across ElasticArcFace and CurricularFace losses, so it is not tied to one architecture or loss.
  • The released subsets give other researchers ready-made training data with characterized fairness and accuracy trade-offs, so new losses or architectures can be tested on top of them.

Reading between the lines

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

  • If the ethnicity score is accepted as a measure of representativeness, the same balancing recipe should transfer to other coarse demographic attributes such as age or gender presentation, where discrete bins are also rough proxies for continuous variation.
  • The identity-level score can be reused as a general data-pruning utility: pairing it with a quality score, which the paper shows to be weakly correlated, could select subsets that are simultaneously fair and high-quality.
  • The paper's own caveat, that the score measures how easily an image is classified into a group, suggests a strong test: replace the classifier score with human-annotated continuous ancestry proportions and see whether the balancing advantage survives; if it does not, the effect is about image difficulty rather than ethnicity.
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 / 6 minor

Summary. The paper argues that ethnicity labels for face-recognition training data should be treated as continuous rather than discrete, and that balancing a training set in this continuous space yields fairer and often more accurate models than balancing by discrete ethnicity counts. The authors construct what they call an ethnicity score by training a four-class softmax ethnicity classifier on top of a fixed ElasticArcFace ResNet-100 feature extractor, using the BUPT-BalancedFace training set itself, and then compute per-image, per-identity, and per-ethnicity scores. They propose three sampling protocols (A, B, C) that iteratively remove low-scoring identities from a selected ethnicity group, plus relabeled variants A(R), B(R), C(R), and compare the resulting models against models trained on the full dataset and on randomly sampled subsets preserving discrete balance. Experiments are carried out with iResNet-34/50/100 and ElasticArcFace/CurricularFace losses, with evaluation on RFW, reporting per-ethnicity accuracy, STD, and SER. The central empirical claim is that continuous-space balancing consistently outperforms discrete-space balancing, with Protocol A giving the largest fairness gains and Protocol B retaining accuracy best. The authors also release code and sampled dataset versions.

Significance. If the central claim holds, the paper would make a useful conceptual contribution: it challenges the common assumption that equalizing discrete demographic counts is sufficient for dataset balance, and it provides a concrete, reproducible recipe for score-guided identity removal that appears to improve fairness metrics without sacrificing overall accuracy. The breadth of the experimental work is a strength: more than 65 trained models, three architectures, two losses, multiple removal fractions, and publicly released code and sampled datasets. The main risk is construct validity: the continuous ethnicity score is produced by a classifier trained on the very dataset being rebalanced, and the paper itself acknowledges (Section 4.4) that a high score means only that the image is easily classified into that group of images in the dataset, not that it is representative in an objective sense. If the score mainly captures in-sample classifiability, typicality, or image utility, then the reported gains could be a data-cleaning artifact rather than evidence about the continuous nature of ethnicity.

major comments (4)
  1. [Section 4.4 and Section 3.1] The ethnicity score is generated by a classifier trained on BUPT-BalancedFace and then applied to the same BUPT-BalancedFace identities to compute IDS and drive the removal protocols. This is an in-sample scoring procedure. The authors themselves state in Section 4.4 that a high probability does not mean the image is representative of the definition of a group, only that it is easily classified as belonging to that group of images in the dataset. Under this definition, the score is a measure of classifiability/typicality relative to the training distribution, and removing the lowest-score identities is equivalent to removing hard or atypical training samples. The reported advantage over random sampling could therefore be explained by data cleaning rather than by any continuous property of ethnicity. A necessary control is an out-of-sample validation of the score: for example, train the ethnicity classifier on a different dataset (e.g., BUPT-GlobalFace or an external ethnically labeled dataset) and apply it to BUPT-BalancedFace, or evaluate on held-out identities to show that the score generalizes as a measure of group representativeness rather than memorization of training identities. Without such a test, the central claim that continuous ethnicity balancing outperforms discrete balancing is not established.
  2. [Section 4.3 and Table 2] The single-ethnicity analysis in Table 2 shows that retaining the highest-scoring African identities gives a 92.78% African accuracy versus 89.95% for the lowest-scoring ones, which is consistent with the paper's assumption. However, this pattern is exactly what any classifier-confidence or utility-based selection would produce: removing the least classifiable identities of a group improves the group's training signal. The comparison with CR-FIQA in Section 4.3 is only a qualitative scatter plot and reports no correlation coefficient; it rules out one quality confound but not pose, blur, identity difficulty, or training-set membership. To separate the ethnicity-representativeness effect from a generic utility effect, the authors should include a control protocol driven by an alternative utility score (e.g., CR-FIQA, image norm, or an unsupervised typicality measure) and show that the ethnicity-guided selection produces different or better fairness outcomes than utility-guided selection at the same removal level. As it stands, Table 2 and the CR-FIQA plot do not distinguish between 'ethnically atypical' and 'hard to classify' as the operative selection axis.
  3. [Section 5.2 and Tables 3-5] All reported results appear to come from a single training run per condition. Many of the claimed improvements are small in absolute terms: for example, at 27k identities in Table 3, Protocol A has STD 1.17 and SER 1.76 versus 1.27 and 1.86 for Random, and in Table 5 Protocol C at 21k has STD 1.60 versus 1.44 for Random, i.e., worse. Without multiple seeds, confidence intervals, or a significance test, the 'consistently outperform' claim in the abstract is not statistically supported. I recommend either reporting results over at least three seeds per condition with means and standard deviations, or conducting a statistical test over bootstrap resamples of the evaluation set, or explicitly stating and discussing the single-seed limitation. This is load-bearing because the central claim is a comparative statement across many conditions, and the magnitude of some differences is comparable to typical training noise.
  4. [Section 5.2.3 and Algorithm 3] Protocol C removes the lowest-scoring identity from the ethnicity group with the highest total ES score, which pushes that group's total score down. The paper describes this as balancing in the continuous space, but the procedure is not symmetric with Protocols A and B: it equalizes the sum rather than the mean, and the paper does not justify why sum-balancing is the right continuous notion of balance for the fairness claim. This matters because Protocol C is reported as having the weakest fairness gains (Table 5). Please clarify the theoretical target of Protocol C and, if it is meant to balance sums, discuss why that target is appropriate; otherwise the comparison among protocols is hard to interpret.
minor comments (6)
  1. [Title and Abstract] There are typos in 'Balacing' and 'continuos' in the title/abstract; these should be corrected.
  2. [Section 4.4] The ethnicity classifier training details are incomplete: no number of epochs, learning rate, batch size, or validation accuracy is reported for the classification layer. Since the entire paper depends on this classifier, these details are necessary for reproducibility.
  3. [Section 5.2.2] The sentence 'these latter strategies seem to hold significant advantages with respect to random sampling' appears in table captions but is not supported by any statistical significance test; please either remove 'significant' or add the appropriate test.
  4. [Section 5.1] The text says 'removing exactly half of the identities in a given ethnicity,' and Table 2 reports 3,500 retained identities, but it is not immediately clear that each ethnicity in BUPT-BalancedFace contains 7,000 identities; please state the group sizes explicitly in the table caption.
  5. [Figure 2 and Figure 6 numbering] The figures are referenced out of order: Figure 6 is referred to in Section 4.4 before Figure 5 is introduced. Please renumber or reorder so that figures appear in the order they are cited.
  6. [Section 4.3] The claim that the ethnicity score is 'uncorrelated' with CR-FIQA is based on visual inspection of a scatter plot; report the actual correlation coefficient (e.g., Pearson and Spearman) for the comparison.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: the 'continuous balance' of the dataset is defined through an in-sample fitted classifier and is achieved by construction, but the central fairness comparison is measured on the external RFW benchmark, so the core claim retains independent content.

  1. self definitional [Section 3.1, Eqs. (4)-(5), Section 3.1.2 Protocol A / Algorithm 1, Figure 3]
    "the identity score highlights the contributions of each identity to their ethnicity score... one can remove the identity j with the lowest IDS (y) Aj from the ethnicity y with the smaller value for diag(ESA)(y). This will lead to an increase in the latter score."

    The 'balance of the dataset' is defined by Eq. (5) as the group mean of fitted identity scores, and Protocol A removes the argmin-score identity from the argmin-score group. Removing the minimum from the minimum always raises that group's mean, so the equalization reported in Figure 3 is the algorithm's objective realized by construction, not an empirical property of an ethnicity continuum. The claim that 'not all identities from one ethnicity contribute equally to the balance of the dataset' is then a tautology of the average-based definition of balance, and the produced subsets are 'balanced in the continuous space' by the same fitted score that defines balance.

  2. fitted input called prediction [Section 4.4 (Ethnicity Classifier) applied in Section 3.1.2-3.1.4 Protocols A-C and Section 3.1.5 Relabeling]
    "After training our classifier, we use it to infer a probability vector for each image in both BUPT-BalancedFace and BUPT-GlobalFace datasets. One important point is to take into account that an image with a high probability on the Caucasian group is not necessarily representative of the definition of Caucasian, it means that that particular image is easily classified as belonging to that group of images in the dataset."

    The pseudo-continuous ethnicity label is the softmax confidence of a classifier whose classification layer is trained on BUPT-BalancedFace and then applied back to the same identities, so the scores are in-sample fitted values. By the paper's own definition, high score means in-sample classifiability rather than objective degree of ethnicity. The protocols 'predict' the least-representative identities as the lowest fitted confidences and delete them; the advantage over random sampling is therefore partly a fitted-confidence pruning effect (removing in-sample-hard, atypical, or mislabelled identities) that the paper attributes to continuous ethnicity.

full rationale

The headline claim—that models trained on datasets balanced in the continuous space outperform models balanced on discrete labels—is tested on the external RFW benchmark with multiple backbones and losses, so the fairness/accuracy comparison itself is not a by-construction identity and gives the paper substantial independent content. Nevertheless, two load-bearing links in the derivation are definitional. First, 'continuous balance' is defined (Eq. 5) as the mean of fitted identity scores, and Protocols A-B remove the argmin-score identity from the argmin-score group, so the equalization in Figure 3 is the algorithm's objective achieved by construction, and 'not all identities contribute equally to the balance' is a tautology of the average definition. Second, the continuous label is the in-sample softmax of a classifier trained on BUPT-BalancedFace and applied back to the same identities; Section 4.4 concedes that a high score only means the image is easily classified, so the selected 'least representative' identities are the fitted-lowest-confidence identities, and the reported gains over random sampling are partly a pruning artifact of in-sample classifiability. The abstract promises 'theoretical validation,' but no theorem is given; that validation reduces to the protocols' definitional convergence. No load-bearing self-citation or imported uniqueness theorem was found—self-citations ([2], [5], [8], [27], [29], [30]) concern losses, evaluation metrics, and related work. The external RFW evaluation and the CR-FIQA comparison limit, but do not remove, the construct-validity circularity, hence the moderate score.

Assumptions & free parameters 2 free parameters · 2 assumptions · 1 invented entities

The central claim rests on a single fitted quantity, the ethnicity score, which is generated by a classifier trained in-sample on the dataset being balanced. The removal counts are hand-chosen experimental conditions. The axioms are domain assumptions about the informativeness of the latent space and an ad hoc definition of what it means to balance in the continuous space. No new physical entities are introduced.

free parameters (2)
  • Ethnicity classifier (ResNet100 + ElasticArc backbone, classification layer trained on BUPT-BalancedFace) = Trained model; weights not enumerated in the paper
    The continuous labels are the softmax outputs of this classifier. The classifier is trained on the same dataset that is then sampled, making the score a fitted quantity. The central results depend on this score.
  • Number of identities removed Z (1k, 3.5k, 7k, 14k) = 1,000 / 3,500 / 7,000 / 14,000
    These removal counts are chosen by hand as 3.5%, 12.5%, 25% and 50% of 28k identities. The paper itself notes they 'might not necessarily reflect the optimal value' for identity removal, and they determine the operating points of all compared protocols.
assumptions (2)
  • domain assumption The latent space of face recognition models encodes ethnicity information, and a classifier trained on top of it yields valid continuous ethnicity scores.
    Section 4.4 states 'We verified that the latent space of FR systems encodes ethnicity information'. This is load-bearing because the entire method uses these scores as ground truth for representativeness.
  • ad hoc to paper Removing identities with the lowest ethnicity score from the group with the lowest average score progressively balances the dataset in a way that improves fairness without sacrificing accuracy.
    This is the design principle behind Algorithms 1-3. It is the paper's redefinition of balance rather than a standard result, and it is validated only indirectly through the downstream accuracy and fairness metrics.
invented entities (1)
  • Continuous ethnicity label (ethnicity flow)
    purpose: Replace discrete ethnicity labels with a continuous degree-of-belonging score per image and per identity.
    Defined as the softmax output of a classifier trained on the paper's own training set. No external ground truth or falsifiable handle is provided; it is a re-interpretation of model confidence. The conclusion mentions 'ethnicity flow' as an introduced concept, but it is not operationalized beyond the classifier score.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Balancing Beyond Discrete Categories: Continuous Demographic Labels for Fair Face Recognition." pith.science (2026). https://pith.science/paper/JRH2A4ED

@misc{pith2026250601532,
  author       = {Pith},
  title        = {Pith review of: Balancing Beyond Discrete Categories: Continuous Demographic Labels for Fair Face Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JRH2A4ED}},
  note         = {Machine review of arXiv:2506.01532}
}
read the original abstract

Bias has been a constant in face recognition models. Over the years, researchers have looked at it from both the model and the data point of view. However, their approach to mitigation of data bias was limited and lacked insight on the real nature of the problem. Here, in this document, we propose to revise our use of ethnicity labels as a continuous variable instead of a discrete value per identity. We validate our formulation both experimentally and theoretically, showcasing that not all identities from one ethnicity contribute equally to the balance of the dataset; thus, having the same number of identities per ethnicity does not represent a balanced dataset. We further show that models trained on datasets balanced in the continuous space consistently outperform models trained on data balanced in the discrete space. We trained more than 65 different models, and created more than 20 subsets of the original datasets.

Figures

Figures reproduced from arXiv: 2506.01532 by the authors.

Figure 1
Figure 1. Demonstration of the Sampling Effect: For each sam￾pling intensity (using sampling strategy A) the sample with the lowest confidence of each label is shown. For instance, on No Sampling, it is possible to see an Asian image (according to the dataset label) that the ethnicity classifier model has a low confi￾dence in the Asian class. It is also noticeable that as the sampling intensifies, the ethnicity traits become … view at source ↗
Figure 2
Figure 2. highlights, in each row, the inter-ethnicity vari￾ability for each of the four distinct demographic groups. On each row, from left to the right one can observe the images with lowest to the highest prevalence of ethnicity specific traits. As shown by the four left images of each row, one can verify not only the existence of possibly noisy labels, but also non-representative samples of each ethnicity and samples that… view at source ↗
Figure 3
Figure 3. The evolution of the different scores diag(ESA) (y) , diag(ESB) (y) , and diag(ESC ) (y) (y-axis) as the number of identities in the dataset decreases (x-axis). It can be seen that for the first two protocols, balance is quickly achieved, whereas for the other only the last iteration is balanced. the largest of the two datasets, with two million images from 38k identities. The ethnicity distribution of these identi￾… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Distribution shift of the ethnicity score on the different [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: For this, we compared our score with the quality score pro￾duced by CR-FIQA [3] [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Identity Score (IDS(y) A ) Distribution for the different demographic groups for both the original ethnicity labels and after a relabelling process. tain the discretely balanced nature of the original set. So, for each sampling strategy, we compare a model trained on i…
Figure 7
Figure 7. Figure 7: Trade-off between bias and performance visually highlighted through the computation of a pareto frontier with all the dominant [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 42 canonical work pages

  1. [1]

    Bacchini and L

    F. Bacchini and L. Lorusso. Race, again: how face recog- nition technology reinforces racial discrimination. Jour- nal of information, communication and ethics in society , 17(3):321–335, 2019

    work page 2019

  2. [2]

    Boutros, N

    F. Boutros, N. Damer, F. Kirchbuchner, and A. Kuijper. Elasticface: Elastic margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Or- leans, LA, USA, June 19-20, 2022, pages 1577–1586. IEEE, 2022

    work page 2022

  3. [3]

    Boutros, M

    F. Boutros, M. Fang, M. Klemt, B. Fu, and N. Damer. CR- FIQA: face image quality assessment by learning sample rel- ative classifiability. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 5836–5845. IEEE, 2023

    work page 2023

  4. [4]

    Buolamwini and T

    J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In S. A. Friedler and C. Wilson, editors, Conference on Fair- ness, Accountability and Transparency, FAT 2018, 23-24 February 2018, New York, NY, USA, volume 81 of Proceed- ings of Machine Learning Research , pages 77–91. PMLR, 2018

    work page 2018

  5. [5]

    MST-KD: Multiple Specialized Teachers Knowledge Distillation for Fair Face Recognition

    E. Caldeira, J. S. Cardoso, A. F. Sequeira, and P. C. Neto. Mst-kd: Multiple specialized teachers knowledge distillation for fair face recognition. arXiv preprint arXiv:2408.16563, 2024

    work page Pith review arXiv 2024

  6. [6]

    Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 13th IEEE International Conference on Automatic Face & Gesture Recognition, FG 2018, Xi’an, China, May 15-19, 2018, pages 67–74. IEEE Computer Society, 2018

    work page 2018

  7. [7]

    P. K. Chandaliya, K. Raja, R. Ramachandra, Z. Akhtar, and C. Busch. Towards inclusive face recognition through syn- thetic ethnicity alteration. arXiv preprint arXiv:2405.01273, 2024

    arXiv 2024

  8. [8]

    DeAndres-Tame, R

    I. DeAndres-Tame, R. Tolosana, P. Melzi, R. Vera- Rodriguez, M. Kim, C. Rathgeb, X. Liu, A. Morales, J. Fier- rez, J. Ortega-Garcia, et al. Frcsyn challenge at cvpr 2024: Face recognition challenge in the era of synthetic data. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3173–3183, 2024

    work page 2024

Show all 44 references
  1. [9]

    DeAndres-Tame, R

    I. DeAndres-Tame, R. Tolosana, P. Melzi, R. Vera- Rodr´ıguez, M. Kim, C. Rathgeb, X. Liu, A. Morales, J. Fi ´errez, J. Ortega-Garcia, Z. Zhong, Y . Huang, Y . Mi, S. Ding, S. Zhou, S. He, L. Fu, H. Cong, R. Zhang, Z. Xiao, E. Smirnov, A. Pimenov, A. Grigorev, D. Timoshenko, K....

    2024

  2. [10]

    J. Deng, J. Guo, J. Yang, N. Xue, I. Kotsia, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recogni- tion. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):5962– 5979, 2022

    2022

  3. [11]

    Dooley, R

    S. Dooley, R. Sukthanker, J. P. Dickerson, C. White, F. Hut- ter, and M. Goldblum. Rethinking bias mitigation: Fairer architectures make for fairer face recognition. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information...

    2023

  4. [12]

    Dooley, G

    S. Dooley, G. Z. Wei, T. Goldstein, and J. Dickerson. Ro- bustness disparities in face detection. In S. Koyejo, S. Mo- hamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems 35: An- nual Conference on Neural Information Pr...

    2022

  5. [13]

    Drozdowski, C

    P. Drozdowski, C. Rathgeb, A. Dantcheva, N. Damer, and C. Busch. Demographic bias in biometrics: A survey on an emerging challenge. IEEE Transactions on Technology and Society, 1(2):89–103, 2020

    2020

  6. [14]

    S. Gong, X. Liu, and A. K. Jain. Jointly de-biasing face recognition and demographic attribute estimation. In A. Vedaldi, H. Bischof, T. Brox, and J. Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX...

    2020

  7. [15]

    S. Gong, X. Liu, and A. K. Jain. Mitigating face recognition bias via group adaptive classifier. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, vir- tual, June 19-25, 2021, pages 3414–3424. Computer Vision Foundation / IEEE, 2021

    2021

  8. [16]

    Y . Guo, L. Zhang, Y . Hu, X. He, and J. Gao. Ms-celeb-1m: Challenge of recognizing one million celebrities in the real world. In J. P. Allebach, Z. Fan, and Q. Lin, editors,Imaging and Multimedia Analytics in a Web and Mobile World 2016, San Francisco, CA, USA, February 14-18...

    2016

  9. [17]

    Gwilliam, S

    M. Gwilliam, S. Hegde, L. Tinubu, and A. Hanson. Re- thinking common assumptions to mitigate racial bias in face recognition datasets. In IEEE/CVF International Confer- ence on Computer Vision Workshops, ICCVW 2021, Mon- treal, BC, Canada, October 11-17, 2021, pages 4106–4115....

    2021

  10. [18]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learn- ing for image recognition. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 770–778, 2016

    2016

  11. [19]

    Huang, M

    L. Huang, M. Wang, J. Liang, W. Deng, H. Shi, D. Wen, Y . Zhang, and J. Zhao. Gradient attention balance network: Mitigating face recognition racial bias via gradient atten- tion. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2023 - Workshops, Vanco...

    2023

  12. [20]

    Huang, Y

    Y . Huang, Y . Wang, Y . Tai, X. Liu, P. Shen, S. Li, J. Li, and F. Huang. Curricularface: Adaptive curriculum learning loss for deep face recognition. In 2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pages 5900–

    2020

  13. [21]

    Huber, M

    M. Huber, M. Fang, F. Boutros, and N. Damer. Are explain- ability tools gender biased? a case study on face presentation attack detection. In 2023 31st European Signal Processing Conference (EUSIPCO), pages 945–949. IEEE, 2023

    2023

  14. [22]

    S. D. Jaiswal, K. Duggirala, A. Dash, and A. Mukherjee. Two-face: Adversarial audit of commercial face recognition systems. In C. Budak, M. Cha, and D. Quercia, editors, Proceedings of the Sixteenth International AAAI Conference on Web and Social Media, ICWSM 2022, Atlanta, Ge...

    2022

  15. [23]

    T. L. Johnson, N. N. Johnson, D. McCurdy, and M. S. Olajide. Facial recognition systems in policing and racial disparities in arrests. Government Information Quarterly , 39(4):101753, 2022

    2022

  16. [24]

    Krishnapriya, V

    K. Krishnapriya, V . Albiero, K. Vangara, M. C. King, and K. W. Bowyer. Issues related to face recognition accuracy varying based on race and skin tone. IEEE Transactions on Technology and Society, 1(1):8–20, 2020

    2020

  17. [25]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In P. L. Bartlett, F. C. N. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Infor- mation Processing Systems 25: 26th Annual C...

    2012

  18. [26]

    E. Levi, T. Xiao, X. Wang, and T. Darrell. Rethinking preventing class-collapsing in metric learning with margin- based losses. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 10296–10305. IEEE, 2021

    2021

  19. [27]

    R. M. Mamede, P. C. Neto, and A. F. Sequeira. Fair- ness under cover: Evaluating the impact of occlusions on demographic bias in facial recognition. arXiv preprint arXiv:2408.10175, 2024

    2024 arXiv

  20. [28]

    Morales, J

    A. Morales, J. Fi ´errez, R. Vera-Rodr´ıguez, and R. Tolosana. Sensitivenets: Learning agnostic representations with appli- cation to face images. IEEE Trans. Pattern Anal. Mach. In- tell., 43(6):2158–2164, 2021

    2021

  21. [29]

    P. C. Neto, E. Caldeira, J. S. Cardoso, and A. F. Sequeira. Compressed models decompress race biases: What quan- tized models forget for fair face recognition. In N. Damer, M. Gomez-Barrero, K. B. Raja, C. Rathgeb, A. F. Sequeira, M. Todisco, and A. Uhl, editors,International ...

    2023

  22. [30]

    P. C. Neto, I. Colakovic, S. Karakati ˇc, and A. F. Sequeira. How knowledge distillation mitigates the synthetic gap in fair face recognition. arXiv preprint arXiv:2408.17399 , 2024

    2024 arXiv

  23. [31]

    P. C. Neto, T. Gonc ¸alves, J. R. Pinto, W. Silva, A. F. Se- queira, A. Ross, and J. S. Cardoso. Causality-inspired tax- onomy for explainable artificial intelligence. arXiv preprint arXiv:2208.09500, 2024

    2024 arXiv

  24. [32]

    I. D. Raji and J. Buolamwini. Actionable auditing revisited: Investigating the impact of publicly naming biased perfor- mance results of commercial AI products. Commun. ACM, 66(1):101–108, 2023

    2023

  25. [33]

    J. P. Robinson, G. Livitz, Y . Henon, C. Qin, Y . Fu, and S. Ti- moner. Face recognition: Too bias, or not too bias? In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR Workshops 2020, Seattle, WA, USA, June 14-19, 2020, pages 1–10. Computer Vision F...

    2020

  26. [34]

    Schroff, D

    F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In IEEE Conference on Computer Vision and Pattern Recogni- tion, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 815–823. IEEE Computer Society, 2015

    2015

  27. [35]

    Serna, A

    I. Serna, A. Morales, J. Fierrez, and N. Obradovich. Sensi- tive loss: Improving accuracy and fairness of face represen- tations with discrimination-aware deep learning. Artificial Intelligence, 305:103682, 2022

    2022

  28. [36]

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265–5274. Compu...

    2018

  29. [37]

    M. Wang, W. Deng, J. Hu, X. Tao, and Y . Huang. Racial faces in the wild: Reducing racial bias by information max- imization adaptation network. In 2019 IEEE/CVF Interna- tional Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 692–

    2019

  30. [38]

    M. Wang, Y . Zhang, and W. Deng. Meta balanced network for fair face recognition. IEEE transactions on pattern anal- ysis and machine intelligence, 44(11):8433–8448, 2021

    2021

  31. [39]

    Z. Wang, K. Qinami, I. C. Karakozis, K. Genova, P. Nair, K. Hata, and O. Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2020, Seattle, WA, USA, June 1...

    2020

  32. [40]

    Wu and K

    H. Wu and K. W. Bowyer. What should be balanced in a ”balanced” face recognition dataset? In 34th British Ma- chine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, page 235. BMV A Press, 2023

    2023

  33. [41]

    Yucer, S

    S. Yucer, S. Akc ¸ay, N. A. Moubayed, and T. P. Breckon. Exploring racial bias within face recognition via per-subject adversarially-enabled data augmentation. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR Workshops 2020, Seattle, WA, USA, June 14...

    2020

  34. [42]

    Z. Zhu, G. Huang, J. Deng, Y . Ye, J. Huang, X. Chen, J. Zhu, T. Yang, D. Du, J. Lu, and J. Zhou. Webface260m: A bench- mark for million-scale deep face recognition. IEEE Trans. Pattern Anal. Mach. Intell., 45(2):2627–2644, 2023. A. Supplementary A.1. Different Architectures I...

    2023

  35. [44]

    However, this strategy achives better fairness metrics than Random sampling

    When compared to our sampling approaches, random sampling is only close, yet with some difference, in perfor- mance to strategy A. However, this strategy achives better fairness metrics than Random sampling. Strategies B and C seem to have lower fairness, but when carefully an...

  36. [5909]

    Computer Vision Foundation / IEEE, 2020

    2020

Pith tools

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