Pith. sign in

REVIEW 3 major objections 4 minor 50 references

Benchmarking Face Recognition without Real Faces

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read Two synthetic face datasets reproduce the ranking behavior of real face-recognition benchmarks across 24 models.

desk verdict First serious attempt at validating synthetic face datasets as real-benchmark substitutes; the core finding is credible, but the model-pool composition needs sensitivity analysis before the confidence intervals can be fully trusted. read the letter →

arxiv 2607.14932 v1 pith:OJ7VEPBL submitted 2026-07-16 cs.CV cs.AI

classification cs.CVcs.AI
keywords facerecognitionsyntheticdatabenchmarkingbiometricevaluationMorphVec2Facecorrelationanalysisprivacy-preserving
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 sets out to prove that evaluating face recognition models on synthetic datasets can stand in for evaluating them on real photographs. The authors compare 12 synthetic datasets against 7 established real benchmarks using 24 pre-trained models, measuring verification accuracy, biometric error rates (EER, FMR100, FMR1000, ZeroFMR), similarity score distributions, and FID-domain distances. They find that two candidates, MorphFace and Vec2Face, agree with real benchmarks about which models are good and which are not, with Spearman's rho above 0.85 and mean Pearson r of 0.90 — agreement within the spread real benchmarks show among themselves. If correct, this closes the last privacy gap in face recognition research: models could be trained and evaluated without any real facial images, only synthetic ones.

What carries the argument

The load-bearing instrument is the cross-dataset correlation analysis: for 24 models, the same verification protocol is run on each synthetic and each real dataset, and the resulting model-level scores (accuracy, EER, ZeroFMR, FMR100, FMR1000) are correlated between every dataset pair. Pearson's r measures linear agreement in absolute scores; Spearman's rho measures whether the ranking of models is preserved. FID scores and mated-versus-non-mated similarity distributions are used to explain why some synthetic datasets succeed (clear identity separability, moderate domain gap) and others fail.

What would settle it

Take a set of recent face recognition models not among the 24 (e.g., newly released transformer architectures from different training regimes) and compute their rankings on MorphFace, Vec2Face, and the seven real benchmarks. If Spearman's rho between the synthetic and real rankings falls below 0.85, or if the 95% confidence intervals no longer overlap the real-real agreement range, the central claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that well-constructed synthetic face datasets can serve as reliable substitutes for real benchmarks in comparative face-recognition evaluation. The evidence is a large cross-correlation study: for each of 24 models, scores on 12 synthetic and 7 real datasets are computed, and agreement between datasets is quantified by Pearson's r and Spearman's rho. MorphFace and Vec2Face consistently achieve mean r = 0.90 and rho = 0.87 across biometric metrics, with 95% confidence intervals overlapping the agreement observed among real datasets themselves. Datasets designed explicitly for benchmarking (ControlFace10k, SynMulti-PIE) are among the worst, while these training-oriented da

Load-bearing premise

The 24-model pool is treated as an unbiased sample of face recognition architectures and training regimes; if it is unrepresentative, the measured agreement between synthetic and real benchmarks could change.

Editorial extensions

If this is right

  • Face recognition models can be benchmarked on synthetic data, removing the need to process or redistribute real biometric photos.
  • The synthetic-benchmark approach enables fully synthetic pipelines, from training to evaluation, lowering legal and ethical exposure under GDPR and BIPA.
  • The validated criteria (r and rho > 0.9 with lower confidence bounds > 0.75) give dataset creators concrete targets for future synthetic datasets.
  • Model rankings produced on MorphFace and Vec2Face are expected to transfer to real photographs, so researchers can trust synthetic comparisons for model selection.
  • The failure of some synthetic datasets warns that not all synthetic data works for benchmarking; identity separability and intra-class diversity are necessary ingredients.

Reading between the lines

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

  • The specific winners, MorphFace and Vec2Face, may be superseded by future generators; the durable contribution is the validation methodology, which could become a standard acceptance test for synthetic benchmarks.
  • Because the model pool includes seven CNNs trained on synthetic data, the correlations might be inflated; recomputing the analysis without those models is a direct robustness test the paper does not report.
  • The same correlation-based validation could be applied to other biometric modalities (fingerprint, iris, voice) where synthetic data is emerging, potentially generalizing the privacy-preserving benchmarking argument.
  • The 0.85–0.90 agreement level may not hold for specialized evaluation niches (extreme pose, cross-age, demographic subgroups) that real benchmarks capture differently; targeted synthetic datasets might be needed.
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

3 major / 4 minor

Summary. The paper asks whether synthetic face datasets can replace real benchmarks for face recognition evaluation. It compares 12 synthetic datasets against 7 established real benchmarks using 24 pre-trained models, computing accuracy, EER, FMR100, FMR1000, ZeroFMR, Pearson/Spearman correlations, FID, and similarity distributions. The main claim is that MorphFace and Vec2Face reproduce the relative ranking behaviour of real benchmarks, reaching agreement levels within the natural disagreement observed among real benchmarks themselves, thereby enabling fully synthetic benchmarking pipelines.

Significance. If the central claim holds, the paper would make a strong contribution to privacy-preserving face recognition: it would allow model evaluation without any real face data. The study is unusually comprehensive in scope: 12 synthetic datasets, 7 real benchmarks, 24 models spanning CNNs and ViTs, and multiple biometric operating points. The public code release and inclusion of the pair lists in the replication package are concrete reproducibility strengths. However, the headline conclusion rests on a small, partially non-independent model pool that includes models trained on the very synthetic datasets being evaluated, and at least one quantitative table is internally inconsistent. These issues are load-bearing and must be resolved before the claims can be accepted.

major comments (3)
  1. [§3.2, Table 2; §4.1–4.5] The 24-model pool includes 7 CNNs trained entirely on synthetic data, and Table 2 indicates that ResNet50 variants trained on Langevin-DisCo, IDiff-Face, HyperFace, and SFace are both pool members and candidate benchmark datasets. A model trained on a synthetic dataset is not an independent probe of that dataset as a benchmark; its score reflects training-distribution match. This can inflate the measured synthetic–real correlations and affect the 'within natural disagreement' claim. Please report a sensitivity analysis that excludes all models trained on the synthetic datasets under evaluation, or otherwise demonstrates that the headline correlations are unchanged.
  2. [§3.2, §3.4, §4.5] The 24 models are not 24 independent observations: the pool contains IResNet18/50/100, EdgeFace base/S/XS/XXS, TransFace S/L, and multiple ResNet50 variants. The bootstrap in §3.4 resamples models with replacement, treating near-duplicates as independent. This makes the 95% CIs narrower than they would be under an appropriate clustered bootstrap (e.g., by architecture family or by training dataset). Since §4.5 uses those CIs to claim overlap with real–real disagreement, the effective sample size should be stated and the CIs recomputed with a cluster-robust procedure.
  3. [Table 6, FMR100 row] For AgeDB-30, the table reports Meanρ=0.81 and MinLowCI=0.87; for CFP-FP it reports Meanρ=0.86 and MinLowCI=0.89. A lower confidence bound cannot exceed the point estimate of the corresponding pairwise correlation. Therefore one of the two quantities is mislabelled, computed on a different set of pairs, or simply erroneous. This table is part of the quantitative evidence for the 'no systematic gap' conclusion in §4.5, so the inconsistency must be fixed and the corrected values checked against the stated conclusion.
minor comments (4)
  1. [§4.3, Table 5] FID is computed with an IResNet18 face-recognition embedding rather than the standard Inception-v3. This is a legitimate choice, but it makes the FID values non-comparable to standard FID numbers in the literature. State this explicitly in the text and justify the choice.
  2. [§3.1, Table 1] The distinction between SFace and SFace2 should be made explicit: SFace is a model and SFace2 is a dataset. The current text can confuse the two, especially in Table 2.
  3. [§4.4, Figure 3] The text says 'SwinFace and IResNet18 produce similar performance levels while FaceNet CASIA clearly underperforms; Vec2Face and MorphFace reproduce this ordering.' Figure 3 presumably shows only a subset of the 24 models; please clarify which models are displayed and whether the statement is about the displayed subset or the full pool.
  4. [General] A minor typographical issue: the first affiliation line contains a stray accented character ('Wendkˆuuni'); the author name should appear as 'Wendkûuni' or similar.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: headline correlations are independently measured against external real benchmarks.

full rationale

The paper's central claim is an empirical benchmark-fidelity comparison: 12 synthetic datasets are scored by 24 pretrained models and correlated (Eqs. 1-3) with 7 established real benchmarks. No parameter is fitted from the real-benchmark scores and then used to predict those same scores; MorphFace and Vec2Face are pre-existing datasets (refs [31,46]), not constructed by this paper. The self-citations to [4] frame dataset-quality notions (intra-class diversity, identity separability) and interpret ZeroFMR outliers, but they are not load-bearing: the rho/r values in Sections 4.1-4.5 are computed from the present measurements, not imported from [4]. The inclusion of 7 synthetic-trained ResNet50 models (Table 2) is a genuine model-pool confound — e.g., an IDiff-Face model scores IDiff-Face — but it is not a definitional circularity: no model in the pool is trained on MorphFace or Vec2Face, and no correlation is forced by construction. One data-integrity issue should be flagged separately: Table 6 reports for FMR100 AgeDB-30 'Mean rho = 0.81' and 'MinLowCI = 0.87', which is impossible if MinLowCI is a minimum lower 95% CI bound; this does not indicate circularity but should be corrected. Overall, the derivation is self-contained against external real benchmarks.

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

No free parameters are fitted and no new entities are introduced. The study relies on domain-level assumptions about the representativeness of the model pool and real benchmarks, plus standard statistical assumptions about correlation metrics and bootstrap intervals.

assumptions (4)
  • domain assumption The seven real benchmarks (LFW, CPLFW, CALFW, AgeDB-30, CFP-FP, IJB-B, IJB-C) provide the correct ground-truth ordering of face recognition models.
    The entire evaluation uses correlation with real benchmarks as the definition of benchmarking fidelity; no external measure is used.
  • domain assumption The 24 pre-trained models are a representative and sufficiently diverse sample of face recognition systems.
    Correlations are computed across these models; if the pool is biased (e.g., over-representing ResNet50 or models trained on the synthetic datasets), the conclusions may not generalize.
  • domain assumption Pearson's r and Spearman's rho between dataset-level model scores are valid proxies for benchmark interchangeability.
    The paper equates high correlation of model ordering with the ability to replace a real benchmark; this ignores calibration and absolute-score differences.
  • domain assumption The Unrestricted with Labeled Outside Data protocol on randomly sampled 3,000 mated and 3,000 non-mated pairs is a fair approximation of benchmarking behavior for synthetic datasets.
    Used in §3.5 to generate synthetic-benchmark scores; if the random pairs are not representative, the resulting correlations would be biased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Face Recognition without Real Faces." pith.science (2026). https://pith.science/paper/OJ7VEPBL

@misc{pith2026260714932,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Face Recognition without Real Faces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJ7VEPBL}},
  note         = {Machine review of arXiv:2607.14932}
}
read the original abstract

Synthetic face datasets have become effective enough to train face recognition models with accuracy rivaling that of models trained on real photographs. This progress sidesteps the ethical and legal burdens of collecting real biometric data, yet evaluation has not kept pace. Even studies that train entirely on synthetic images still rely on real-face benchmarks to measure performance, leaving the privacy problem only half solved. We ask whether synthetic datasets can replace real benchmarks for face recognition evaluation. We test 12 synthetic datasets against 7 established real benchmarks using 24 pre-trained models that span both convolutional and transformer architectures. Our evaluation covers biometric verification metrics, similarity score distributions, cross-model ranking consistency, and the underlying distributional properties of each dataset. Benchmarking fidelity varies widely across the synthetic candidates, but the two strongest, MorphFace and Vec2Face, reproduce the relative behavior of real benchmarks and reach agreement levels that fall within the natural disagreement already observed among the real benchmarks themselves. These results establish that well-constructed synthetic datasets can support reliable comparative evaluation for face recognition, moving the field closer to a fully synthetic and privacy-preserving pipeline for both training and benchmarking.

Figures

Figures reproduced from arXiv: 2607.14932 by the authors.

Figure 1
Figure 1. Recognition accuracy alignment between datasets across 24 models - Pearson’s [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Spearman’s ρ with 95% CI represents how strongly 24 models’ accuracies ranking is preserved across datasets [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Mated vs. non-mated similarity distributions comparison across individual models and datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Per-metric Spearman’s ρ on 24 models – MorphFace and Vec2Face vs. real datasets. two datasets are nearly indistinguishable from each other in MAE, confirming that either can serve as a benchmark. 5. Discussion Our experiments distinguish reliable from unreliable synthe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 7 linked inside Pith

  1. [1]

    X. An, X. Zhu, Y . Xiao, L. Wu, M. Zhang, Y . Gao, B. Qin, D. Zhang, and Y . Fu. Partial fc: Training 10 million identities on a single machine, 2021

  2. [2]

    Baltsou, I

    G. Baltsou, I. Sarridis, C. Koutlis, and S. Papadopoulos. Sdfd: Building a versatile synthetic face image dataset with diverse attributes. In2024 IEEE 18th International Con- ference on Automatic Face and Gesture Recognition (FG), pages 1–10. IEEE, 2024

  3. [3]

    Behrmann, W

    J. Behrmann, W. Grathwohl, R. T. Chen, D. Duvenaud, and J.-H. Jacobsen. Invertible residual networks. InInternational conference on machine learning, pages 573–582. PMLR, 2019

  4. [4]

    Borsukiewicz, F

    P. Borsukiewicz, F. Boutros, I. E. Olatunji, C. Beumier, W. C. Ou´edraogo, J. Klein, and T. F. Bissyand´e. Beyond real faces: synthetic datasets can achieve reliable recognition perfor- mance without privacy compromise.npj Artificial Intelli- gence, 2026

  5. [5]

    Boutros, J

    F. Boutros, J. H. Grebe, A. Kuijper, and N. Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650–19661, 2023

  6. [6]

    Boutros, M

    F. Boutros, M. Huber, A. T. Luu, P. Siebke, and N. Damer. Sface2: Synthetic-based face recognition with w-space identity-driven sampling.IEEE Transactions on Biometrics, Behavior, and Identity Science, 2024

  7. [7]

    Boutros, M

    F. Boutros, M. Huber, P. Siebke, T. Rieber, and N. Damer. Sface: Privacy-friendly and accurate face recognition using synthetic data. In2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1–11. IEEE, 2022

  8. [8]

    Boutros, V

    F. Boutros, V . Struc, J. Fierrez, and N. Damer. Synthetic data for face recognition: Current state and future prospects. Image and Vision Computing, 135:104688, 2023

Show all 50 references
  1. [9]

    Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018

  2. [10]

    Colbois, T

    L. Colbois, T. de Freitas Pereira, and S. Marcel. On the use of automatically generated synthetic image datasets for bench- marking face recognition. In2021 IEEE International Joint Conference on Biometrics (IJCB), pages 1–8. IEEE, 2021

  3. [11]

    J. Dan, Y . Liu, H. Xie, J. Deng, H. Xie, X. Xie, and B. Sun. Transface: Calibrating transformer training for face recog- nition from a data-centric perspective. InProceedings of the IEEE/CVF international conference on computer vision, pages 20642–20653, 2023

  4. [12]

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019

  5. [13]

    J. Deng, J. Guo, D. Zhang, Y . Deng, X. Lu, and S. Shi. Lightweight face recognition challenge. In2019 IEEE/CVF International Conference on Computer Vision Workshop (IC- CVW), pages 2638–2646, 2019

  6. [14]

    Regulation (EU) 2016/679 of the European Parliament and of the Council

    European Parliament and Council of the European Union. Regulation (EU) 2016/679 of the European Parliament and of the Council

  7. [15]

    Geissb ¨uhler, H

    D. Geissb ¨uhler, H. O. Shahreza, and S. Marcel. Syn- thetic face datasets generation via latent space explo- ration from brownian identity diffusion.arXiv preprint arXiv:2405.00228, 2024

  8. [16]

    George, C

    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, 6(2):158–168, 2024

  9. [17]

    George and S

    A. George and S. Marcel. Digi2real: Bridging the realism gap in synthetic data face recognition via foundation models. InProceedings of the Winter Conference on Applications of Computer Vision, pages 1469–1478, 2025

  10. [18]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Gen- erative adversarial nets.Advances in neural information pro- cessing systems, 27, 2014

  11. [19]

    G. E. Granato. Statistics for stochastic modeling of volume reduction, hydrograph extension, and water-quality treat- ment by structural stormwater runoff best management prac- tices (bmps). Technical report, US Geological Survey, 2014

  12. [20]

    Grother, P

    P. Grother, P. Grother, A. Hom, M. Ngan, and K. Hanaoka. Face recognition vendor test (FRVT) Part 7: Identification for paperless travel and immigration. US Department of Commerce, National Institute of Standards and Technology, 2021

  13. [21]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

  14. [22]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion proba- bilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

  15. [23]

    G. B. Huang, M. Mattar, T. Berg, and E. Learned-Miller. La- beled faces in the wild: A database forstudying face recog- nition in unconstrained environments. InWorkshop on faces in’Real-Life’Images: detection, alignment, and recognition, 2008

  16. [24]

    Public act 095-0994, 2008

    Illinois General Assembly. Public act 095-0994, 2008

  17. [25]

    Information technology – biometric performance testing and reporting – part 1: Prin- ciples and framework, 2021

    ISO/IEC JTC 1/SC 37 Biometrics. Information technology – biometric performance testing and reporting – part 1: Prin- ciples and framework, 2021

  18. [26]

    Karras, M

    T. Karras, M. Aittala, S. Laine, E. H ¨ark¨onen, J. Hellsten, J. Lehtinen, and T. Aila. Alias-free generative adversarial networks.Advances in neural information processing sys- tems, 34:852–863, 2021

  19. [27]

    M. Kim, F. Liu, A. Jain, and X. Liu. Dcface: Synthetic face generation with dual condition diffusion model. InProceed- ings of the ieee/cvf conference on computer vision and pat- tern recognition, pages 12715–12725, 2023

  20. [28]

    Kortylewski, A

    A. Kortylewski, A. Schneider, T. Gerig, B. Egger, A. Morel- Forster, and T. Vetter. Training deep face recognition systems with synthetic data.arXiv preprint arXiv:1802.05891, 2018

  21. [29]

    B. Maze, J. Adams, J. A. Duncan, N. Kalka, T. Miller, C. Otto, A. K. Jain, W. T. Niggel, J. Anderson, J. Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In2018 international conference on biometrics (ICB), pages 158–165. IEEE, 2018

  22. [30]

    Melzi, C

    P. Melzi, C. Rathgeb, R. Tolosana, R. Vera-Rodriguez, D. Lawatsch, F. Domin, and M. Schaubert. Gandiffface: Controllable generation of synthetic datasets for face recog- nition with realistic variations. InProceedings of the IEEE/CVF International Conference on Computer Vision...

  23. [31]

    Y . Mi, Z. Zhong, Y . Huang, Q. Yuan, X. Zhao, J. Xu, S. Ding, S. Wang, R. Guo, and S. Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21203–21214, 2025

  24. [32]

    Moschoglou, A

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kot- sia, and S. Zafeiriou. Agedb: the first manually collected, in-the-wild age database. Inproceedings of the IEEE confer- ence on computer vision and pattern recognition workshops, pages 51–59, 2017

  25. [33]

    Nisevic, D

    M. Nisevic, D. Milojevic, and D. Spajic. Synthetic data in medicine: Legal and ethical considerations for patient profil- ing.Computational and Structural Biotechnology Journal, 28:190–198, 2025

  26. [34]

    Nzalasse, R

    K. Nzalasse, R. Raj, E. Laird, and C. Clark. Sig: A syn- thetic identity generation pipeline for generating evaluation datasets for face recognition. InInternational Conference on Pattern Recognition, pages 299–313. Springer, 2024

  27. [35]

    L. Qin, M. Wang, C. Deng, K. Wang, X. Chen, J. Hu, and W. Deng. Swinface: A multi-task transformer for face recog- nition, expression recognition, age estimation and attribute estimation.IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2223–2234, 2023

  28. [36]

    H. Qiu, B. Yu, D. Gong, Z. Li, W. Liu, and D. Tao. Syn- face: Face recognition with synthetic data. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 10880–10890, 2021

  29. [37]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer. High-resolution image synthesis with latent diffu- sion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684– 10695, 2022

  30. [38]

    Sengupta, J.-C

    S. Sengupta, J.-C. Chen, C. Castillo, V . M. Patel, R. Chel- lappa, and D. W. Jacobs. Frontal to profile face verification in the wild. In2016 IEEE winter conference on applications of computer vision (WACV), pages 1–9. IEEE, 2016

  31. [39]

    H. O. Shahreza and S. Marcel. Hyperface: Generating syn- thetic face recognition datasets by exploring face embedding hypersphere.arXiv preprint arXiv:2411.08470, 2024

  32. [40]

    Smith and S

    M. Smith and S. Miller. The ethical application of biometric facial recognition technology.Ai & Society, 37(1):167–175, 2022

  33. [41]

    Sun and Z

    Z. Sun and Z. Liu. Ensuring privacy in face recognition: a survey on data generation, inference and storage.Discover Applied Sciences, 7(5):441, 2025

  34. [42]

    Z. Sun, S. Song, I. Patras, and G. Tzimiropoulos. Cemi- face: Center-based semi-hard synthetic face generation for face recognition.Advances in Neural Information Process- ing Systems, 37:35612–35638, 2024

  35. [43]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9, 2015

  36. [44]

    Szegedy, V

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

  37. [45]

    Whitelam, E

    C. Whitelam, E. Taborsky, A. Blanton, B. Maze, J. Adams, T. Miller, N. Kalka, A. K. Jain, J. A. Duncan, K. Allen, et al. Iarpa janus benchmark-b face dataset. Inproceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90–98, 2017

  38. [46]

    H. Wu, J. Singh, S. Tian, L. Zheng, and K. W. Bowyer. Vec2face: Scaling face dataset generation with loosely con- strained vectors.arXiv preprint arXiv:2409.02979, 2024

  39. [47]

    Yeung, T

    M. Yeung, T. Teramoto, S. Wu, T. Fujiwara, K. Suzuki, and T. Kojima. Variface: Fair and diverse synthetic dataset gener- ation for face recognition.arXiv preprint arXiv:2412.06235, 2024

  40. [48]

    D. Yi, Z. Lei, S. Liao, and S. Z. Li. Learning face represen- tation from scratch.arXiv preprint arXiv:1411.7923, 2014

  41. [49]

    Zheng and W

    T. Zheng and W. Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained envi- ronments.Beijing University of Posts and Telecommunica- tions, Tech. Rep, 5(7):5, 2018

  42. [50]

    Zheng, W

    T. Zheng, W. Deng, and J. Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained en- vironments.arXiv preprint arXiv:1708.08197, 2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.