Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Investigation of Accuracy and Bias in Face Recognition Trained with Synthetic Data

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Demographically balanced synthetic face data—particularly from Stable Diffusion v3.5—can reduce racial bias in face-recognition models, while synthetic-only training still trails real-data training on the hardest generalization benchmarks.

desk verdict Useful scale-controlled benchmark for synthetic face data, but the images-per-identity claim is confounded with optimizer updates and the bias finding needs replication. read the letter →

arxiv 2507.20782 v1 pith:Z3UN4WLD submitted 2025-07-28 cs.CV cs.AI

Pavel Korshunov , Ketan Kotwal , Christophe Ecabert , Vidit Vidit , Amir Mohammadi , Sebastien Marcel This is my paper
classification cs.CVcs.AI
keywords syntheticfacedatademographicbiasrecognitionfairnessFairGenStableDiffusionv3.5identityaugmentationIJB-B/CevaluationRacialFacesintheWild
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 asks whether synthetic face images can train face-recognition models that are both accurate and fair across racial groups. The authors build FairFaceGen, a demographically balanced synthetic dataset, by prompting two text-to-image generators (Flux.1-dev and Stable Diffusion v3.5) across gender, age, and race, then expanding each identity with several identity-preserving augmentation methods. Training the same recognition model on this data and on size-matched real and synthetic baselines, they find that balanced synthetic data—especially SD35-based data—can lower the spread of accuracy across racial groups on the Racial Faces in the Wild benchmark, even below real-data baselines. However, models trained only on synthetic data still generalize worse on the demanding IJB-B/C benchmarks than models trained on real data. The paper also shows that the number of images per identity and the choice of augmentation generator shift both accuracy and fairness, making those generation choices practical design levers.

What carries the argument

The central mechanism is FairFaceGen, a prompt-driven generation pipeline: balanced text-to-image seed generation with SD35 or Flux produces distinct identities, and Arc2Face or four IP-Adapter variants produce intra-class variations, all evaluated under a size-controlled protocol that fixes roughly 10K identities across every training set while varying images per identity. A supporting instrument is the cosine-distance analysis of EdgeFace embeddings, which measures how well augmentations preserve an identity (intra-class) and how separated different identities are (inter-class); the paper uses these distances to explain why SD35 supports fairness while Flux, with lower inter-class separation, supports accuracy less well.

What would settle it

Train the same AdaFace model on an SD35-generated dataset with identical prompts, augmentation methods, and images-per-identity, but with identity demographics sampled in real-world proportions instead of balanced, and compare the RFW standard deviation; if the unbalanced version shows the same low spread, then balance itself is not the active ingredient.

Watch

Extended reading notes

Core claim

The central claim is that demographic balancing during synthetic generation is a working lever for fairness: the SD35-based balanced sets produced the lowest standard deviation of accuracy across RFW's four racial groups (1.06 for SD35-All-IPA, versus 3.03 for a size-matched real WebFace42M subset), meaning the fairer behavior does not require giving up on synthetic data. At the same time, the paper finds that synthetic data alone remains behind real data on IJB-B/C, where the best synthetic dataset (DCFace-10K) trails WebFace42M-10K by about 10% in true-accept rate; it also documents that increasing images per identity improves standard-benchmark accuracy but can degrade IJB-B/C performance, and that mixing many augmentation generators hurts generalization. The paper interprets these results as evidence that seed-generator identity separation and augmentation quality—not just dataset scale—drive the accuracy and fairness of synthetic-data-trained models.

Load-bearing premise

The claim that demographic balancing causes the fairness gain rests on the assumption that the lower spread of accuracy across racial groups is due to the balanced composition of the generated images, and not to some other property of the SD35 generator, the augmentation recipe, or the training setup.

Editorial extensions

If this is right

  • Balanced synthetic data gives practitioners a prompt-based way to reduce racial accuracy gaps without training a custom generator.
  • Because SD35-based balanced data reaches lower RFW bias than real-data baselines, fairness gains are achievable even when no balanced real dataset is available.
  • The images-per-identity curves give concrete guidance: roughly 16 augmentations per identity gives the best IJB-B/C generalization, while 24-32 can improve standard benchmarks but hurt hard-benchmark performance.
  • The near-zero IJB-B/C result for the randomly subsampled DCFace-10K-Small shows that standard-benchmark accuracy can conceal catastrophic generalization failure, so hard benchmarks should be part of synthetic-data evaluation.
  • Hybrid training that mixes balanced synthetic data with real data is identified by the authors as the promising next step toward combining high accuracy with low bias.

Reading between the lines

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

  • A natural extension the authors did not run: fine-tune a real-data-trained model on balanced SD35 data and measure whether the RFW fairness gain survives without an IJB-B/C accuracy penalty.
  • Since training epochs and batch size are fixed while images per identity varies, datasets with more images per identity receive more optimization steps per identity; holding total updates constant would separate the data-composition effect from the training-budget effect.
  • The paper measures bias as the standard deviation of accuracy across four RFW groups at one decision threshold; a stricter test would report per-group true-accept rates at fixed false-accept rates on IJB-B/C, which would show whether fairness holds in high-security operating points.
  • Because Flux images look more like polished studio portraits while SD35 images look more like in-the-wild photos, some of SD35's fairness advantage may come from image style rather than demographic balance; varying style while keeping balance fixed would isolate that factor.
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. The paper investigates whether face recognition models trained on demographically balanced synthetic datasets can achieve competitive accuracy and reduced racial bias. The authors generate a balanced dataset, FairFaceGen, using two text-to-image generators (Flux.1-dev and Stable Diffusion v3.5), augment identities with Arc2Face and four IP-Adapter variants, and train a ResNet-50 AdaFace model. They compare against size-matched real and synthetic baselines on standard benchmarks, IJB-B/C, and RFW, reporting that synthetic data still lags real data on IJB-B/C but that SD35-based balanced synthetic data shows lower racial-bias standard deviation on RFW. They also report that the number of images per identity affects accuracy and fairness.

Significance. If the reported effects are robust, the paper provides a useful controlled benchmark: it equalizes identity count across datasets, includes IJB-B/C evaluations that are often missing in synthetic-data face recognition studies, compares multiple augmentation generators under a fixed backbone, and makes a practical attempt at prompt-based demographic balancing. The dataset and protocol could serve as a reference for future work on fair synthetic face recognition. However, the central causal claims are currently supported only by single training runs without error bars, and the images-per-identity manipulation is confounded with the number of optimizer updates. These issues limit the strength of the conclusions as they stand.

major comments (4)
  1. [§4.1 and §5.1] The claim that the number of images per identity drives accuracy and fairness is confounded with the number of optimizer updates. Training is fixed at 30 epochs with batch size 256 (§4.1), so for roughly 11K identities, 8 images per identity yields about 10.3K SGD updates while 32 images per identity yields about 41.3K updates. The monotonic gains on AgeDB30/CFP-FP and the nonmonotonic IJB-B/C pattern in Table 1 could therefore reflect training compute or overfitting rather than data composition. The same confound affects the cross-dataset comparisons in Tables 1 and 2, since WebFace42M-10K, CASIA-WebFace, DCFace, and the FairFaceGen variants differ in images per identity and thus in the number of training steps. I ask the authors to add a control that matches the number of optimizer steps or otherwise demonstrate that the observed effects are not explained by training duration.
  2. [Tables 1 and 2] All experimental results come from single training runs with no error bars, repeats, or statistical testing. This is especially concerning given the DCFace-10K-Small result in Table 1, which shows a near-total collapse on IJB-B/C (TAR 0.1% at FAR 1e-4) while DCFace-10K achieves 77–89% despite the same identities and only a reduction in images per identity. The authors declare this behavior 'out of scope,' but it is central to their images-per-identity argument and to the reliability of the benchmark. Without repeated runs or additional diagnostics (e.g., training curves, loss trajectories, or multiple seeds), the patterns in the tables—including the RFW standard deviations that ground the bias-mitigation claim—may be noise. Please provide repeated runs for the central comparisons and report means with variance or confidence intervals.
  3. [§5.2 and Conclusion] The bias-mitigation claim rests on the standard deviation of subgroup accuracies on RFW (Table 2). The reported differences, e.g., SD35-All-IPA with Std 1.06 versus Digi2Real with 2.39 or WebFace42M-10K with 3.03, are not accompanied by any uncertainty estimate, and the group-level accuracies for SD35 are all between roughly 74% and 77%. Moreover, the decision threshold is fit on the complete RFW set, which is a standard evaluation step, but it means the reported values are already optimized on the test set; combined with single runs, a 1–2 point standard-deviation difference is not established as reliable. The conclusion that demographic balancing 'significantly reduce[s] racial bias' should be softened or supported by statistical evidence.
  4. [§5.1, DCFace-10K-Small] The unexplained cliff between DCFace-10K-Small and DCFace-10K is not a peripheral anomaly: it appears in the main results table and is used to motivate the importance of images per identity. The text states that 'a dedicated evaluation of the DCFace would be required to better understand this behavior, but such investigation is out of the scope of this paper.' Since the central claim concerns the effect of images per identity, this behavior cannot simply be set aside. At minimum, the authors should either exclude DCFace-10K-Small from the causal interpretation and justify why, or provide a diagnostic analysis showing whether the collapse is due to training instability, overfitting, or a data-composition effect.
minor comments (5)
  1. [§4] In the sentence introducing the evaluation, 'IJC-B/C datasets' should be 'IJB-B/C datasets.'
  2. [§5.2] The text refers to 'Dig2Real' but the dataset is named 'Digi2Real' elsewhere; please correct the typo.
  3. [Figure 4] The histograms in Figure 4 lack axis labels and a clear legend for the datasets shown; adding these would help readers interpret the inter- and intra-class distance claims.
  4. [§3.1] The authors state that exactly the same prompts were used for Flux and SD35, but no prompt templates or example prompts are provided; including them would improve reproducibility.
  5. [Table 1] The naming of DCFace-10K is confusing: the text in §5.1 calls it the 'full DCFace 0.5M' while the table header says 10K. Please clarify the relationship between DCFace-10K, DCFace-10K-Small, and the original DCFace dataset.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: an empirical benchmark whose claims rest on external test sets; self-citations are non-load-bearing.

full rationale

The paper's derivation chain is: construct FairFaceGen with controlled seed and augmentation generators (Section 3.1), train a fixed AdaFace/ResNet-50 pipeline (Section 4.1), and evaluate on external benchmarks LFW, AgeDB-30, CFP-FP, CP-LFW, CA-LFW, IJB-B/C, and RFW (Section 3.3, Tables 1-2). The central claims—that SD35-based balanced synthetic data reduces RFW standard deviation and that the number and quality of intra-class augmentations affect accuracy and fairness—are empirical observations obtained from those external test sets. No parameter is fitted to the target claim and then reported as a prediction; the RFW decision threshold is set on the full RFW set as a standard evaluation step, not as a construction of the claimed bias-mitigation effect. Self-citations appear (EdgeFace for embedding analysis and Digi2Real/HyperFace as baselines), but these are evaluation instruments or comparison points, and the headline conclusions are benchmarked against independent external data (WebFace42M-10K, CASIA, IJB-B/C, and RFW). No uniqueness theorem or prior same-author result is invoked to force the choice of SD35 over alternatives. The most serious validity caveat—varying images-per-identity while fixing epochs and batch size also varies the number of optimizer updates (Section 4.1 versus Section 5.1)—is a confound affecting causal interpretation, not a circular reduction: Table 1 values are not equal to the generation parameters by construction. The acknowledged 'out of scope' DCFace-10K-Small anomaly is a stated limitation, not a self-referential step. Therefore the derivation is self-contained and no circularity is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical quantities. Its claims rest on empirical measurements, with the main free parameter being the RFW decision threshold.

free parameters (1)
  • RFW decision threshold = Optimized on the full RFW test set
    The paper selects the threshold that maximizes accuracy on the complete RFW dataset and applies it to each racial subgroup; the reported group accuracies and standard deviation depend on this chosen value.
assumptions (3)
  • domain assumption Text-to-image generators produce identifiable face identities whose demographic attributes can be controlled via prompts.
    Section 3.1 assumes that varying prompt attributes yields balanced identity counts and that the generators create distinct identities suitable for FR training.
  • domain assumption The EdgeFace pretrained model provides a valid measure of inter/intra-class similarity for comparing synthetic and real face distributions.
    Section 3.4 uses EdgeFace embeddings to assess dataset quality; EdgeFace was trained on real faces, so it may not be a neutral yardstick for synthetic data.
  • domain assumption The RFW dataset is an appropriate benchmark for racial bias in face recognition, and single-threshold evaluation on it gives stable conclusions.
    Section 4.2 uses RFW to measure bias without repeated cross-validation or statistical testing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigation of Accuracy and Bias in Face Recognition Trained with Synthetic Data." pith.science (2026). https://pith.science/paper/Z3UN4WLD

@misc{pith2026250720782,
  author       = {Pith},
  title        = {Pith review of: Investigation of Accuracy and Bias in Face Recognition Trained with Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3UN4WLD}},
  note         = {Machine review of arXiv:2507.20782}
}
read the original abstract

Synthetic data has emerged as a promising alternative for training face recognition (FR) models, offering advantages in scalability, privacy compliance, and potential for bias mitigation. However, critical questions remain on whether both high accuracy and fairness can be achieved with synthetic data. In this work, we evaluate the impact of synthetic data on bias and performance of FR systems. We generate balanced face dataset, FairFaceGen, using two state of the art text-to-image generators, Flux.1-dev and Stable Diffusion v3.5 (SD35), and combine them with several identity augmentation methods, including Arc2Face and four IP-Adapters. By maintaining equal identity count across synthetic and real datasets, we ensure fair comparisons when evaluating FR performance on standard (LFW, AgeDB-30, etc.) and challenging IJB-B/C benchmarks and FR bias on Racial Faces in-the-Wild (RFW) dataset. Our results demonstrate that although synthetic data still lags behind the real datasets in the generalization on IJB-B/C, demographically balanced synthetic datasets, especially those generated with SD35, show potential for bias mitigation. We also observe that the number and quality of intra-class augmentations significantly affect FR accuracy and fairness. These findings provide practical guidelines for constructing fairer FR systems using synthetic data.

Figures

Figures reproduced from arXiv: 2507.20782 by the authors.

Figure 1
Figure 1. Examples of images cropped to 112 × 112 from WebFace42M [2] and synthetic datasets Digi2Real [3], HyperFace [4], DCFace [5], and generated by us with SD35 and Flux. different and the total number of images are different, which are the parameters affecting the performance of the model. Such mismatches conceal the true effect of generation parameters [9] and a fair evaluation needs to acknowledge the issue and equaliz… view at source ↗
Figure 2
Figure 2. Generation process of FairFaceGen dataset. relying on the embeddings from either InsightFace [18] or CLIP [19], pretrained on LAION-2B [20] and COYO-700M [21] datasets. 4. Measure inter-identity and intra-identity similarity with an independent EdgeFace [22] face recognition model and compare different variations of the synthetic data with HyperFace [4], Digi2Real [3] and a size-matched 10K identities subset from th… view at source ↗
Figure 3
Figure 3. Seed generators SD35 (top of (a)) and Flux (bottom of (a)) and their augmentations generated with Arc2Face [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Mean intra-class and min inter-class cosine distances for synthetic datasets and WebFace-10K. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: ROC computed on IJB-B and IJB-C databases for AdaFace models trained on faces generated with 16, 24, or [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages

  1. [1]

    GANDiffFace: Control- lable generation of synthetic datasets for face recognition with gans and diffusion models,

    P. Melzi, C. Rathgeb, R. Tolosana, R. Vera-Rodriguez, D. Lawatsch, F. Domin, and M. Schaubert, “GANDiffFace: Control- lable generation of synthetic datasets for face recognition with gans and diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2023. 1

    work page 2023

  2. [2]

    WebFace260M: A benchmark unveiling the power of million-scale deep face recognition,

    Z. Zhu, G. Huang, J. Deng, Y . Ye, J. Huang, X. Chen, J. Zhu, T. Yang, J. Lu, D. Du, et al., “WebFace260M: A benchmark unveiling the power of million-scale deep face recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10492–10502, 2021. 1, 2, 3, 4, 5, 7

    work page 2021

  3. [3]

    Digi2Real: Bridging the realism gap in synthetic-data face recognition via foundation models,

    A. George and S. Marcel, “Digi2Real: Bridging the realism gap in synthetic-data face recognition via foundation models,” in Proc. WACV Workshops, 2025. 1, 2, 3, 4, 5, 7

    work page 2025

  4. [4]

    HyperFace: Generating synthetic face-recognition datasets by exploring the face-embedding hypersphere,

    H. O. Shahreza and S. Marcel, “HyperFace: Generating synthetic face-recognition datasets by exploring the face-embedding hypersphere,” in Accepted in International Conference on Learning Representations (ICLR) , 2025. 1, 2, 3, 5, 7

    work page 2025

  5. [5]

    DCFace: Synthetic face generation with dual condition diffusion model,

    M. Kim, F. Liu, A. K. Jain, and X. Liu, “DCFace: Synthetic face generation with dual condition diffusion model,” in Proc. CVPR, 2023. 1, 2, 3, 5, 7

    work page 2023

  6. [6]

    Idiff-face: Synthetic-based face recognition through fizzy identity- conditioned diffusion models,

    F. Boutros, J. H. Grebe, A. Kuijper, and N. Damer, “Idiff-face: Synthetic-based face recognition through fizzy identity- conditioned diffusion models,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pp. 19593–19604,

    work page 2023

  7. [7]

    Synthetic face datasets generation via latent space exploration from brownian identity diffusion,

    D. Geissb ¨uhler, H. O. Shahreza, and S. Marcel, “Synthetic face datasets generation via latent space exploration from brownian identity diffusion,” in Accepted in International Conference on Machine Learning (ICML) , 2025. 1

    work page 2025

  8. [8]

    Variface: Fair and diverse synthetic dataset generation for face recognition,

    M. Yeung, T. Teramoto, S. Wu, T. Fujiwara, K. Suzuki, and T. Kojima, “Variface: Fair and diverse synthetic dataset generation for face recognition,” 2025. 1, 3, 4

    work page 2025

Show all 41 references
  1. [9]

    The impact of balancing real and synthetic data on accuracy and fairness in face recognition,

    A. Atzori, P. Cosseddu, G. Fenu, and M. Marras, “The impact of balancing real and synthetic data on accuracy and fairness in face recognition,” arXiv preprint arXiv:2409.02867, 2024. 2

    2024 arXiv

  2. [10]

    Vulnerability of automatic identity recognition to audio-visual deep- fakes,

    P. Korshunov, H. Chen, P. N. Garner, and S. Marcel, “Vulnerability of automatic identity recognition to audio-visual deep- fakes,” in 2023 IEEE International Joint Conference on Biometrics (IJCB) , pp. 1–10, 2023. 2

    2023

  3. [11]

    Bias and diversity in synthetic-based face recognition,

    M. Huber, A. T. Luu, F. Boutros, A. Kuijper, and N. Damer, “Bias and diversity in synthetic-based face recognition,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024. 2, 4

    2024

  4. [12]

    From fake to real: Pretraining on balanced synthetic images to prevent spurious correlations in image recognition,

    M. Qraitem, K. Saenko, and B. A. Plummer, “From fake to real: Pretraining on balanced synthetic images to prevent spurious correlations in image recognition,” in Proc. ECCV, 2024. Extended version: arXiv:2308.04553. 2, 3

    2024 arXiv

  5. [13]

    Review of demographic fairness in face recognition,

    K. Kotwal and S. Marcel, “Review of demographic fairness in face recognition,” arXiv preprint arXiv:2502.02309, 2025. 2, 3

    2025 arXiv

  6. [14]

    1.58-bit flux,

    C. Yang, C. Liu, X. Deng, D. Kim, X. Mei, X. Shen, and L.-C. Chen, “1.58-bit flux,” arXiv preprint arXiv:2412.18653, 2024. 2, 5

    2024 arXiv

  7. [15]

    Stable diffusion 3: Scaling rectified flow transformers for high-resolution image synthesis,

    S. AI, “Stable diffusion 3: Scaling rectified flow transformers for high-resolution image synthesis,” arXiv preprint arXiv:2403.03206, 2024. 2, 5

    2024 arXiv

  8. [16]

    Arc2Face: A foundation model for id- consistent human faces,

    F. P. Papantoniou, A. Lattas, S. Moschoglou, J. Deng, B. Kainz, and S. Zafeiriou, “Arc2Face: A foundation model for id- consistent human faces,” arXiv preprint arXiv:2403.11641, 2024. 2, 3, 5

    2024 arXiv

  9. [17]

    IP-Adapter: Text-compatible image prompt adapter for diffusion models,

    L. Zhang, H. Qiu, and J. Liao, “IP-Adapter: Text-compatible image prompt adapter for diffusion models,” arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5

    2023 arXiv

  10. [18]

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

    J. Deng, J. Guo, E. Ververas, I. Kotsia, and S. Zafeiriou, “Retinaface: Single-shot multi-level face localisation in the wild,” in CVPR, 2020. 3, 5

    2020

  11. [19]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. , “Learning transferable visual models from natural language supervision,” in International conference on machine learning , pp. 8748–8763, PmLR, 2021. 3, 5

    2021

  12. [20]

    Laion-5b: An open large-scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al., “Laion-5b: An open large-scale dataset for training next generation image-text models,”Advances in neural information processing systems, vol. 35, pp. ...

    2022

  13. [21]

    Coyo-700m: Image-text pair dataset

    M. Byeon, B. Park, H. Kim, S. Lee, W. Baek, and S. Kim, “Coyo-700m: Image-text pair dataset.” https://github. com/kakaobrain/coyo-dataset, 2022. 3

    2022

  14. [22]

    Edgeface: Efficient face recognition model for edge devices,

    A. George, C. Ecabert, H. O. Shahreza, K. Kotwal, and S. Marcel, “Edgeface: Efficient face recognition model for edge devices,” IEEE Transactions on Biometrics, Behavior , and Identity Science, vol. 6, no. 2, pp. 158–168, 2024. 3, 6 10

    2024

  15. [23]

    Labeled faces in the wild: A database for studying face recognition in unconstrained environments,

    G. B. Huang, M. Mattar, T. Berg, and E. Learned-Miller, “Labeled faces in the wild: A database for studying face recognition in unconstrained environments,” in Workshop on Faces in ’Real-Life’ Images: Detection, Alignment, and Recognition , 2008. 3, 4, 5

    2008

  16. [24]

    Agedb: The first manually collected, in-the-wild age database,

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia, and S. Zafeiriou, “Agedb: The first manually collected, in-the-wild age database,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 51–59, 2017. 3, 4, 5

    2017

  17. [25]

    IARPA Janus Benchmark — B (IJB-B): Face recognition benchmarking,

    C. Whitelam, E. Taborsky, A. Beveridge, et al., “IARPA Janus Benchmark — B (IJB-B): Face recognition benchmarking,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , p. 90–98, 2017. 3, 4, 5, 7

    2017

  18. [26]

    IARPA Janus Benchmark — C: Face recognition in video,

    B. Maze, J. Adams, J. Duncan, et al., “IARPA Janus Benchmark — C: Face recognition in video,” in IEEE Internationalt Conference on Biometrics (ICB), pp. 158–165, 2018. 3, 4, 5, 7

    2018

  19. [27]

    Racial faces in the wild: Reducing racial bias by information-maximization adaptation network,

    M. Wang, W. Deng, J. Hu, X. Tao, and Y . Huang, “Racial faces in the wild: Reducing racial bias by information-maximization adaptation network,” in Proc. ICCV, 2019. 3, 6, 7, 9

    2019

  20. [28]

    Learning face representation from scratch,

    D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Learning face representation from scratch,” arXiv preprint arXiv:1411.7923, 2014. 3, 5, 7

    2014 arXiv

  21. [29]

    Analyzing and improving the image quality of styleGAN,

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of styleGAN,” in Proc. CVPR, pp. 8107–8116, 2020. 3

    2020

  22. [30]

    Demographic fairness transformer for bias mitigation in face recognition,

    K. Kotwal and S. Marcel, “Demographic fairness transformer for bias mitigation in face recognition,” in 2024 IEEE Interna- tional Joint Conference on Biometrics (IJCB) , pp. 1–10, IEEE, 2024. 3, 7

    2024

  23. [31]

    Synthetic data for the mitigation of demographic biases in face recognition,

    P. Melzi, C. Rathgeb, R. Tolosana, R. Vera-Rodr´ıguez, A. Morales, D. Lawatsch, F. Domin, and M. Schaubert, “Synthetic data for the mitigation of demographic biases in face recognition,” in Proc. IJCB (Special Session: Synthetic Data in Biometrics) ,

  24. [32]

    Frcsyn challenge at cvpr 2024: Face recognition challenge in the era of synthetic data,

    I. DeAndres-Tame, R. Tolosana, P. Melzi, R. Vera-Rodriguez, M. Kim, C. Rathgeb, X. Liu, A. Morales, J. Fierrez, 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 Com...

    2024

  25. [33]

    FRCSyn challenge at W ACV 2024: Face recognition challenge in the era of synthetic data,

    P. Melzi, R. Tolosana, R. Vera-Rodriguez, M. Kim, C. Rathgeb, X. Liu, I. DeAndres-Tame, A. Morales, J. Fierrez, J. Ortega- Garcia, et al., “FRCSyn challenge at W ACV 2024: Face recognition challenge in the era of synthetic data,” inProceedings of the IEEE/CVF Winter Conference...

    2024

  26. [34]

    FRCSyn-onGoing: Benchmarking and comprehensive evaluation of real and synthetic data to improve face recognition systems,

    P. Melzi, R. Tolosana, R. Vera-Rodriguez, M. Kim, C. Rathgeb, X. Liu, I. DeAndres-Tame, A. Morales, J. Fierrez, J. Ortega- Garcia, et al., “FRCSyn-onGoing: Benchmarking and comprehensive evaluation of real and synthetic data to improve face recognition systems,” Information Fu...

    2024

  27. [35]

    AdaFace: Quality adaptive margin for face recognition,

    M. Kim, A. K. Jain, and X. Liu, “AdaFace: Quality adaptive margin for face recognition,” in Proc. CVPR, 2022. 4, 6

    2022

  28. [36]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 4690–4699, 2019. 5

    2019

  29. [37]

    Cross-age lfw: A database for studying cross-age face recognition in unconstrained environ- ments,

    T. Zheng, W. Deng, and J. Hu, “Cross-age lfw: A database for studying cross-age face recognition in unconstrained environ- ments,” arXiv preprint arXiv:1708.08197, 2017. 5

    2017 arXiv

  30. [38]

    Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments,

    T. Zheng and W. Deng, “Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments,” Tech. Rep. 18-01, Beijing University of Posts and Telecommunications, 2018. 5

    2018

  31. [39]

    Frontal to profile face verification in the wild,

    S. Sengupta, J.-C. Chen, C. Castillo, V . M. Patel, R. Chellappa, and D. W. Jacobs, “Frontal to profile face verification in the wild,” in 2016 IEEE Winter Conference on Applications of Computer Vision (WACV) , pp. 1–9, IEEE, 2016. 5

    2016

  32. [40]

    Joint face detection and alignment using multitask cascaded convolutional networks,

    K. Zhang, Z. Zhang, Z. Li, and Y . Qiao, “Joint face detection and alignment using multitask cascaded convolutional networks,” IEEE signal processing letters, vol. 23, no. 10, pp. 1499–1503, 2016. 6

    2016

  33. [41]

    Mitigating demographic bias in face recognition via regularized score calibration,

    K. Kotwal and S. Marcel, “Mitigating demographic bias in face recognition via regularized score calibration,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pp. 1150–1159, 2024. 7 11

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.