A Deep Learning Approach for Facial Attribute Manipulation and Reconstruction in Surveillance and Reconnaissance

Anees Nashath Shaik; Barbara Villarini; Vasileios Argyriou

REVIEW 4 major objections 5 minor 19 references

A Deep Learning Approach for Facial Attribute Manipulation and Reconstruction in Surveillance and Reconnaissance

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a three-stage GAN pipeline that generates synthetic skin tones, removes eyeglasses, and enhances image quality can rebalance surveillance face datasets and reduce recognition bias, with the evidence being high PSNR…

desk verdict A workmanlike integration of existing GAN components whose central fairness/recognition claim is not supported by the experiments reported. read the letter →

arxiv 2506.06578 v1 pith:UXZLJESS submitted 2025-06-06 cs.CV

Anees Nashath Shaik , Barbara Villarini , Vasileios Argyriou This is my paper

classification cs.CV

keywords facerecognitionbiasgenerativeadversarialnetworkssynthetictrainingdataWGAN-GPeyeglassesremovalimageenhancementCelebAsurveillance

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that synthetic data generation can compensate for the biases and quality gaps that make surveillance face recognition unreliable, particularly for darker skin tones and faces with glasses. It proposes a pipeline that first analyzes a dataset for underrepresented attributes, then uses a Wasserstein GAN to synthesize skin-tone variation, an encoder-decoder GAN to remove eyeglasses while preserving identity, and a style-transfer enhancement module to clean up degraded footage. These generated and enhanced images are combined with the original data into a supposedly more balanced training set. A sympathetic reader would care because the approach promises a data-side fix for algorithmic fairness in face recognition without requiring new real-world data collection. The paper's own support is limited to image-quality metrics on CelebA outputs, not on downstream recognition accuracy.

What carries the argument

The load-bearing machinery is the adversarial generation stack: WGAN-GP supplies stable training for skin-tone modification through a gradient-penalty-enforced Lipschitz critic; ERGAN uses an encoder-decoder with skip connections and an attention module to remove glasses while an added face-recognition loss and PatchGAN discriminator push for identity preservation; AnimeGANv3 contributes style-transfer enhancement through a double-tail generator and Linearly Adaptive Denormalization. Together they are meant to turn a biased, low-quality dataset into a high-quality balanced dataset for downstream recognition.

What would settle it

Train a face recognizer on the original CelebA split and on the augmented split, then measure recognition accuracy separately for dark-skin-tone subjects and eyeglasses wearers; if accuracy or the between-group accuracy gap does not improve despite the reported PSNR and SSIM gains, the paper's central claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the proposed platform improves both training-data diversity and model fairness, leading to significant gains in recognition performance for underrepresented demographic groups and challenging facial conditions. The mechanism is a three-stage deep-learning pipeline: a WGAN-GP skin-tone modification model that alters pigmentation while keeping identity, an ERGAN eyeglasses-removal model with skip connections, an attention module, a face-recognition loss, and a PatchGAN discriminator, and an AnimeGANv3-based enhancement module using dual output tails and linearly adaptive denormalization. Evaluated on CelebA, the authors report a PSNR of about 29.8-30.0 across the three modules, an SSIM of 0.934 for skin-tone generation, 0.385 for eyeglasses removal, and 0.786 for enhancement, and they interpret these numbers as evidence of high fidelity, preserved identity, and reduced bias in facial recognition.

Load-bearing premise

The load-bearing assumption is that high PSNR and SSIM on generated images reliably predict improved face-recognition accuracy and fairness, something the paper never verifies with a recognition model or a per-group accuracy measurement.

Editorial extensions

If this is right

If the pipeline works as claimed, surveillance face datasets can be rebalanced synthetically instead of collecting new photos of underrepresented groups.
Recognition accuracy should improve specifically for darker skin tones and for faces with glasses, shrinking the performance gaps these groups experience.
The enhancement module should make low-resolution, noisy, or occluded surveillance frames more usable for downstream face analysis.
The same attribute-manipulation architecture could be redirected at other underrepresented attributes, such as masks, expressions, or aging, to further diversify training data.

Reading between the lines

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

The paper's evidence stops at image-quality metrics; the next step would be to run an actual face recognizer on original versus augmented training sets and report accuracy by skin-tone and glasses subgroups, which would directly test the fairness claim.
Since CelebA contains curated celebrity photos, the transfer to real surveillance footage is untested; a domain-shift evaluation on CCTV-style imagery would clarify whether the enhancement and augmentation gains survive in deployment conditions.
The eyeglasses module removes a common occlusion, but the symmetric operation of adding synthetic occlusions to faces could be a low-cost way to further balance training data for masked or partially covered faces.

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.

Desk Editor's Note

read the letter

The paper assembles three existing generative components—WGAN-GP for skin-tone modification, an encoder-decoder GAN with attention for eyeglass removal, and AnimeGANv3 for style-based enhancement—into a single pipeline aimed at balancing surveillance face datasets. The integration itself is the only new element, and that is a modest but legitimate contribution. The writing is clear, the WGAN-GP derivation is standard and correct, and the authors honestly report the low SSIM (0.385) for eyeglass removal, attributing it to the substantial structural change the task requires. That transparency deserves credit.

The soft spot is load-bearing. The abstract and conclusion claim that the platform "enhances both training data diversity and model fairness," leading to "significant improvement in recognition performance" for underrepresented groups. The results section, however, contains only PSNR and SSIM on generated images. No face recognizer is trained or evaluated, no accuracy is broken down by skin tone or occlusion, no fairness metric is computed, and there is no baseline comparison against the original dataset. A high PSNR/SSIM on synthetic images does not imply that adding them to a training set improves recognition or reduces bias. In fact, synthetic augmentation can sometimes worsen worst-group performance, so the claim is not just unverified; it is the kind of claim that needs a specific downstream experiment. The paper also asserts identity preservation in the skin-tone model and the eyeglass removal model, but provides no identity-retention verification. Minor issues include a typo in the tables and a somewhat odd use of AnimeGANv3, which is designed for artistic style transfer, as a surveillance enhancement module without validation of that adaptation.

Who is this for? Researchers working on GAN-based data augmentation for face recognition might find the pipeline description useful as a starting point, and the paper could be a reasonable workshop or system-demo submission. As a research contribution, the current evaluation does not support the central claims. That said, the gap is fixable: the authors need to run a downstream face recognition experiment with fairness metrics and compare against the unaugmented dataset. Given the topic's relevance and the clarity of the methodology, I would send it to peer review rather than desk-reject, but I would expect the reviewers to demand that missing experiment before acceptance.

Referee Report

4 major / 5 minor

Summary. The paper proposes a three-stage data-generation pipeline intended to improve fairness and recognition performance in surveillance face recognition: a WGAN-GP model for skin tone modification, an encoder-decoder GAN (ERGAN) with attention and identity loss for eyeglasses removal, and an AnimeGANv3-based image enhancement module. The authors evaluate the three modules on CelebA using PSNR and SSIM, reporting high PSNR values and a notably low SSIM of 0.385 for eyeglasses removal. The central claim, repeated in the abstract and Sections I and V, is that the platform enhances training data diversity and model fairness and significantly improves recognition performance for underrepresented groups, but this claim is not tested by the reported experiments.

Significance. The problem addressed—bias and data imbalance in face recognition, particularly for dark skin tones and occluded faces—is important and timely. The proposed pipeline is clearly described at the architectural level, and the use of WGAN-GP with gradient penalty and a PatchGAN-based removal model follows standard practice. I do not see a circularity problem: the losses are conventional and no fitted quantity is presented as a prediction. However, the current contribution is only a pipeline proposal with image-quality statistics; there are no downstream recognition experiments, no fairness metrics, and no baselines. Even if the generated images look good, the main claimed benefit is untested. If the missing experiments were added and supported the claims, this could be a useful contribution to fairness-oriented data augmentation for face recognition; at present the significance is not established.

major comments (4)

[Abstract; Section I; Section IV, Tables I-II] The central claim, stated in the Abstract and Section I, is that the platform enhances training data diversity and model fairness and leads to significant improvements in recognition performance for underrepresented demographic groups. Section IV and Tables I and II, however, report only PSNR and SSIM on generated images. No face recognition model is trained on the augmented data, no accuracy is reported by skin-tone group or occlusion status, and no fairness metric such as demographic parity or equalized odds is computed. PSNR and SSIM are image-reconstruction quality scores and do not imply that the generated images improve downstream recognition or reduce bias; synthetic augmentation can even degrade worst-group accuracy. This missing evaluation directly undermines the central claim.
[Section IV, Tables I-II] The quantitative evaluation contains no baseline comparisons. The PSNR and SSIM values for skin tone modification, eyeglasses removal, and enhancement are not compared with any existing method (for example, StarGAN, CycleGAN, SRGAN, or the AnimeGANv3 baseline), and the augmented dataset is not compared with the original CelebA dataset in any downstream task. Without such comparisons, the reported numbers are uncontextualized and the claimed improvements are not established.
[Section III.2; Section IV, Table II] The paper states that the eyeglasses removal model preserves identity through a face recognition loss, but no loss equation is given and no identity-similarity metric is measured. The reported SSIM of 0.385 for eyeglasses removal indicates a large structural change, so identity preservation cannot be inferred from the image-quality metrics. The authors should report face embedding similarity (for example, cosine similarity in a pre-trained face recognition embedding space) between input and output faces, in addition to recognition results on the augmented data.
[Section III.C; Section III.A] The 'dataset bias analysis' that is supposed to detect missing skin tones and glasses is not described with any concrete output, and the claim that the final dataset is 'high-quality balanced' is not verified by any attribute distribution or demographic distribution before and after augmentation. In addition, Section III.A says that domain adaptation techniques are applied to align AnimeGANv3 with surveillance footage, but no details are provided; because AnimeGANv3 is designed for artistic anime-style photo animation, this module could introduce a style shift that harms recognition. Both points need experimental validation.

minor comments (5)

[Throughout] There are typos, including 'Ehnanced' in Tables I and II, 'represenatation' in Section III.2, and a corrupted figure caption 'figureERGAN architecture...' following Figure 3; these should be corrected.
[Section III.C] The dataset split is stated as 70% training, 10% evaluation, and 20% testing; it would be clearer to specify which split is used for training the GANs and which is used for the reported evaluation metrics, since the metrics are computed on generated images.
[Section IV] The text mentions a 'qualitative analysis' but shows no qualitative examples; a figure with sample input/output pairs for each module would help the reader assess the claimed visual quality.
[Section III] Hyperparameters for ERGAN and AnimeGANv3 (learning rates, number of epochs, loss weights, including the weight of the face recognition loss) are not reported; these details are needed for reproducibility.
[Section II; References] The Section II heading is typeset as 'RELATEDWORK' and reference [16] contains extra spaces in its title; these are formatting issues that need cleanup.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency: the reported PSNR/SSIM evaluations do not reduce to the fairness claim by construction, and self-citations are not load-bearing.

full rationale

The paper's central claim—that the platform improves recognition performance and fairness for underrepresented groups—is not derived from a fitted quantity later renamed as a prediction. The pipeline uses standard WGAN-GP, an encoder-decoder GAN with PatchGAN and an identity loss, and AnimeGANv3 as an enhancement module. Losses in Eqs. (1)-(6) are standard adversarial objectives; PSNR and SSIM in Tables I and II are measured on generated images and are not used as substitutes for the recognition or fairness outcome in any equation or statistical procedure. No theorem or parameter is imported from prior work by the same authors to force the choice of architecture; self-citations [1], [2], [5], [17], [18] are background references. The main weakness is evidentiary, not circular: the abstract and Section IV assert that image quality implies reduced bias and better recognition, but no face recognizer, demographic subgroup accuracy, or fairness metric is reported. That is an unsupported inference rather than a derivation that is equivalent to its own inputs. Per the rubric, an unsupported claim without a self-referential reduction does not constitute circularity.

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

The central claim rests on four unverified domain assumptions, and training hyperparameters are reported without sensitivity analysis. No new physical entities are introduced; ERGAN is a named model, not a postulated entity.

free parameters (3)

WGAN-GP critic and generator learning rates = alpha_D = alpha_G = 1e-4
Chosen in Section III without sensitivity analysis; they affect training success but are not the fairness result.
training batch size = 64
Reported in Section III; no ablation or justification is given.
input image resolution = 128x128 (AnimeGANv3 at 256x256)
Determines model capacity and could affect output quality; no ablation is presented.

assumptions (4)

domain assumption CelebA attribute labels represent surveillance-relevant demographic and occlusion distributions.
The pipeline is trained on CelebA and transferred to surveillance without domain shift analysis (Sections III-C and IV).
domain assumption Identity is preserved by skin tone editing and glasses removal.
Only qualitative discussion in Section IV; no identity matching or recognition metric after manipulation is reported.
domain assumption AnimeGANv3 style transfer enhances recognition-relevant details rather than destroying them.
Applied in Section III-A without evaluating recognition on enhanced frames.
ad hoc to paper PSNR and SSIM proxy for recognition fairness.
Tables I and II are the only quantitative evidence offered for the fairness and recognition claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Deep Learning Approach for Facial Attribute Manipulation and Reconstruction in Surveillance and Reconnaissance." pith.science (2026). https://pith.science/paper/UXZLJESS

@misc{pith2026250606578,
  author       = {Pith},
  title        = {Pith review of: A Deep Learning Approach for Facial Attribute Manipulation and Reconstruction in Surveillance and Reconnaissance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UXZLJESS}},
  note         = {Machine review of arXiv:2506.06578}
}

read the original abstract

Surveillance systems play a critical role in security and reconnaissance, but their performance is often compromised by low-quality images and videos, leading to reduced accuracy in face recognition. Additionally, existing AI-based facial analysis models suffer from biases related to skin tone variations and partially occluded faces, further limiting their effectiveness in diverse real-world scenarios. These challenges are the results of data limitations and imbalances, where available training datasets lack sufficient diversity, resulting in unfair and unreliable facial recognition performance. To address these issues, we propose a data-driven platform that enhances surveillance capabilities by generating synthetic training data tailored to compensate for dataset biases. Our approach leverages deep learning-based facial attribute manipulation and reconstruction using autoencoders and Generative Adversarial Networks (GANs) to create diverse and high-quality facial datasets. Additionally, our system integrates an image enhancement module, improving the clarity of low-resolution or occluded faces in surveillance footage. We evaluate our approach using the CelebA dataset, demonstrating that the proposed platform enhances both training data diversity and model fairness. This work contributes to reducing bias in AI-based facial analysis and improving surveillance accuracy in challenging environments, leading to fairer and more reliable security applications.

Figures

Figures reproduced from arXiv: 2506.06578 by the authors.

**Figure 2.** Skin model architecture Input Face (with Glasses) Conv + ReLU Conv + ReLU Conv + ReLU Latent Space (Bottleneck) Deconv + ReLU Deconv + ReLU Deconv + ReLU Output Face (Without Glasses) Attention Module Face Recognition Loss PatchGAN Discriminator figureERGAN architecture with encoder-decoder, skip connections, attention module, and identity loss. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** ERGAN architecture with encoder-decoder, skip connections, attention [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Real-World Photo Generator Encoder (Conv + LADE) Residual Blocks + Attention Double Output Tails (Main + Support) Main Tail (Refined Output) Support Tail (Coarse Output) Discriminator D1 (Refined Image Evaluation) Discriminator D2 (Coarse Image Evaluation) Final Anim…

**Figure 4.** Figure 4: Overview of the AnimeGANv3 architecture, illustrating the genera [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

[1]

Learning how to analyse crowd behaviour using synthetic data,

A. R. Khadka, M. M. Oghaz, W. Matta, M. Cosentino, P. Remagnino, and V . Argyriou, “Learning how to analyse crowd behaviour using synthetic data,” ser. CASA ’19. New York, NY , USA: Association for Computing Machinery, 2019, p. 11–14

work page 2019
[2]

Photometric stereo with an arbitrary number of illuminants,

V . Argyriou, M. Petrou, and S. Barsky, “Photometric stereo with an arbitrary number of illuminants,”Computer Vision and Image Under- standing, vol. 114, no. 8, pp. 887–900, 2010

work page 2010
[3]

Face image quality assessment: A literature survey,

T. Schlett, C. Rathgeb, O. Henniger, J. Galbally, J. Fierrez, and C. Busch, “Face image quality assessment: A literature survey,”ACM Comput. Surv., vol. 54, no. 10s, Sep. 2022. [Online]. Available: https://doi.org/10.1145/3507901

doi:10.1145/3507901 2022
[4]

Issues related to face recognition accuracy varying based on race and skin tone,

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, vol. 1, no. 1, pp. 8–20, 2020

work page 2020
[5]

An ai- assisted skincare routine recommendation system in xr,

G. M. Rajegowda, Y . Spyridis, B. Villarini, and V . Argyriou, “An ai- assisted skincare routine recommendation system in xr,” inAI Technolo- gies and Virtual Reality, K. Nakamatsu, S. Patnaik, and R. Kountchev, Eds. Singapore: Springer Nature Singapore, 2024, pp. 381–395

work page 2024
[6]

Deep learning face attributes in the wild,

Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” inProceedings of International Conference on Computer Vision (ICCV), December 2015

2015
[7]

Investigating bias and fairness in facial expression recognition,

T. Xu, J. White, S. Kalkan, and H. Gunes, “Investigating bias and fairness in facial expression recognition,” inComputer Vision–ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16. Springer, 2020, pp. 506–523

work page 2020
[8]

Bias and fairness in face detection,

H. F. Menezes, A. S. Ferreira, E. T. Pereira, and H. M. Gomes, “Bias and fairness in face detection,” in2021 34th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI). IEEE, 2021, pp. 247–254

work page 2021

Show all 19 references

[9]

A deep dive into dataset imbalance and bias in face identification,

V . Cherepanova, S. Reich, S. Dooley, H. Souri, J. Dickerson, M. Gold- blum, and T. Goldstein, “A deep dive into dataset imbalance and bias in face identification,” inProceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 2023, pp. 229–247

2023
[10]

A comprehensive study on face recognition biases beyond demographics,

P. Terh ¨orst, J. N. Kolf, M. Huber, F. Kirchbuchner, N. Damer, A. M. Moreno, J. Fierrez, and A. Kuijper, “A comprehensive study on face recognition biases beyond demographics,”IEEE Transactions on Tech- nology and Society, vol. 3, no. 1, pp. 16–30, 2021

2021
[11]

Human face super-resolution on poor quality surveillance video footage,

M. Farooq, M. N. Dailey, A. Mahmood, J. Moonrinta, and M. Ekpa- nyapong, “Human face super-resolution on poor quality surveillance video footage,”Neural Computing and Applications, vol. 33, pp. 13 505–13 523, 2021

2021
[12]

Empirical analysis of denoising algo- rithms for cctv face images,

S. R. Bhagwat and L. Ragha, “Empirical analysis of denoising algo- rithms for cctv face images,” inInternational Conference on Artificial Intelligence on Textile and Apparel. Springer, 2023, pp. 255–266

2023
[13]

Low resolution face recognition system based on esrgan,

C. Song, Z. He, Y . Yu, and Z. Zhang, “Low resolution face recognition system based on esrgan,” in2021 3rd International Conference on Applied Machine Learning (ICAML). IEEE, 2021, pp. 76–79

2021
[14]

A novel double-tail generative adversarial network for fast photo animation,

G. LIU, X. CHEN, and Z. GAO, “A novel double-tail generative adversarial network for fast photo animation,”IEICE Transactions on Information and Systems, vol. E107.D, no. 1, pp. 72–82, 2024

2024
[15]

Image-to-image translation using pix2pix gan and cycle gan,

R. K. Senapati, R. Satvika, A. Anmandla, G. Ashesh Reddy, and C. Anil Kumar, “Image-to-image translation using pix2pix gan and cycle gan,” inInternational Conference on Data Intelligence and Cognitive Informatics. Springer, 2023, pp. 573–586

2023
[16]

A Style-Based Generator Architecture for Generative Adversarial Networks ,

T. Karras, S. Laine, and T. Aila, “ A Style-Based Generator Architecture for Generative Adversarial Networks ,”IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 43, no. 12, pp. 4217–4228, Dec. 2021

2021
[17]

Stylemask: Disentangling the style space of stylegan2 for neural face reenactment,

S. Bounareli, C. Tzelepis, V . Argyriou, I. Patras, and G. Tzimiropoulos, “Stylemask: Disentangling the style space of stylegan2 for neural face reenactment,” in2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), 2023, pp. 1–8

2023
[18]

Finding directions in gan’s latent space for neural face reenactment,

S. Bounareli, V . Argyriou, and G. Tzimiropoulos, “Finding directions in gan’s latent space for neural face reenactment,”British Machine Vision Conference (BMVC), 2022

2022
[19]

Improved training of wasserstein gans,

I. Gulrajani, F. Ahmed, M. Arjovsky, V . Dumoulin, and A. C. Courville, “Improved training of wasserstein gans,”Advances in neural informa- tion processing systems, vol. 30, 2017

2017

Pith tools

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

[1] [1]

Learning how to analyse crowd behaviour using synthetic data,

A. R. Khadka, M. M. Oghaz, W. Matta, M. Cosentino, P. Remagnino, and V . Argyriou, “Learning how to analyse crowd behaviour using synthetic data,” ser. CASA ’19. New York, NY , USA: Association for Computing Machinery, 2019, p. 11–14

work page 2019

[2] [2]

Photometric stereo with an arbitrary number of illuminants,

V . Argyriou, M. Petrou, and S. Barsky, “Photometric stereo with an arbitrary number of illuminants,”Computer Vision and Image Under- standing, vol. 114, no. 8, pp. 887–900, 2010

work page 2010

[3] [3]

Face image quality assessment: A literature survey,

T. Schlett, C. Rathgeb, O. Henniger, J. Galbally, J. Fierrez, and C. Busch, “Face image quality assessment: A literature survey,”ACM Comput. Surv., vol. 54, no. 10s, Sep. 2022. [Online]. Available: https://doi.org/10.1145/3507901

doi:10.1145/3507901 2022

[4] [4]

Issues related to face recognition accuracy varying based on race and skin tone,

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, vol. 1, no. 1, pp. 8–20, 2020

work page 2020

[5] [5]

An ai- assisted skincare routine recommendation system in xr,

G. M. Rajegowda, Y . Spyridis, B. Villarini, and V . Argyriou, “An ai- assisted skincare routine recommendation system in xr,” inAI Technolo- gies and Virtual Reality, K. Nakamatsu, S. Patnaik, and R. Kountchev, Eds. Singapore: Springer Nature Singapore, 2024, pp. 381–395

work page 2024

[6] [6]

Deep learning face attributes in the wild,

Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” inProceedings of International Conference on Computer Vision (ICCV), December 2015

2015

[7] [7]

Investigating bias and fairness in facial expression recognition,

T. Xu, J. White, S. Kalkan, and H. Gunes, “Investigating bias and fairness in facial expression recognition,” inComputer Vision–ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16. Springer, 2020, pp. 506–523

work page 2020

[8] [8]

Bias and fairness in face detection,

H. F. Menezes, A. S. Ferreira, E. T. Pereira, and H. M. Gomes, “Bias and fairness in face detection,” in2021 34th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI). IEEE, 2021, pp. 247–254

work page 2021