Pith. sign in

REVIEW 3 major objections 6 minor 35 references

This paper argues that facial retrieval in degraded, high-stakes settings can replace fixed top-k lists with per-query adaptive set sizes that come with a formal guarantee: the true match is included with probability at least 1−α, at confid

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-02 07:50 UTC pith:4TMDJI7B

load-bearing objection A competent applied extension of RCIR to face retrieval, with a real gap between the OOD claims and the actual calibration protocol on SCFace. the 3 major comments →

arxiv 2607.16279 v1 pith:4TMDJI7B submitted 2026-07-09 cs.CV

A Step Forward Towards Trustworthy Risk-Aware Facial Retrieval (RA-FR)

classification cs.CV
keywords facial retrievalrisk controlconformal predictionuncertainty quantificationadaptive retrieval setblind face restorationvision transformersurveillance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

RA-FR claims that facial retrieval in degraded, high-stakes settings can replace fixed top-k lists with retrieval sets whose size adapts per query, and that this set carries a formal guarantee: with user-chosen risk α and confidence 1−δ, the true match is included. The paper argues the guarantee holds because risk is controlled by calibrating a single global scale against a held-out set using Hoeffding's inequality, then scaling each query's set size by its estimated uncertainty. On IMFDB it reports meeting a 5% risk target with an average retrieval set of about 10 images, and reports that stronger embeddings plus blind restoration shrink sets further under domain shift.

Core claim

The central claim is that a risk-aware controller can certify retrieval coverage at deployment time: after calibrating one parameter κ on an i.i.d. calibration set, the online adapter returns for each query X the top ⌈κ·Φ[f_u(X)]⌉ matches, where Φ maps an uncertainty score to a scale. Because the loss—missing the true match—is monotone decreasing in set size, the paper uses Hoeffding's inequality to bound finite-sample risk and picks the smallest κ whose upper confidence bound stays below α. Deploying this κ guarantees ρ(κ) ≤ α with probability 1−δ, turning retrieval into an auditable decision rule. The authors further claim that combining blind restoration (DiffBIR plus an LCM refinement) w

What carries the argument

The risk-controlling threshold κ: a single positive scalar calibrated offline. For a query, the retrieval set size is K(X,κ)=⌈κ·Φ[f_u(X)]⌉, with Φ a normalizing map of the uncertainty head f_u. The empirical miscoverage risk ρ̂(κ) plus the Hoeffding upper-confidence term sqrt(ln(1/δ)/(2n)) is kept below α, and monotonicity of risk in κ lets the controller pick the tightest valid κ by grid search. This is what converts uncertainty scores into a certified set size.

Load-bearing premise

The guarantee depends on the calibration set being drawn from the same distribution as the queries seen at deployment; if the deployment distribution shifts (e.g., new cameras, resolutions, or demographics), the calibrated κ may no longer keep the true-match risk below α.

What would settle it

Calibrate κ on the IMFDB training/calibration split at α=0.05 and δ=0.1, then run the exact same adapter on a held-out SCFace probe set without recalibration. If the empirical fraction of queries whose true match is missing exceeds 0.05 (or breaches the stated confidence margin), the deployment-time guarantee as stated does not hold under shift. The paper does not report this cross-calibration number; reporting it would settle whether the guarantee transfers.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fixed top-k heuristics can be replaced by per-query set sizes with a stated worst-case risk of missing the true match.
  • A user can set α and δ up front; the system will report the implied average set size, making the reliability-efficiency trade-off explicit and auditable.
  • The guarantee is finite-sample and distribution-free with respect to the calibration set, so it does not rely on model assumptions about the feature space.
  • Improvements in embedding quality or input restoration translate directly into smaller certified sets at the same risk level.
  • The framework applies beyond faces to any retrieval task with a monotone miscoverage loss and an uncertainty signal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Implicit in the paper: the guarantee is marginal over the calibration distribution, not conditional on each query; per-query reliability would need a stronger notion such as conditional coverage.
  • The reported SCFace results are presented as out-of-distribution evidence, but the calibration procedure described uses a held-out set; if SCFace is used only as test data without recalibration, the empirical risk on SCFace is informative but is not covered by the Hoeffding guarantee. A direct test would be to calibrate on IMFDB and measure miscoverage on SCFace.
  • A natural extension is to feed the estimated uncertainty back into the restoration step or to gate retrieval with a lightweight OOD detector that enlarges κ when shift is detected; the paper names this as future work.
  • The same risk-control mechanism could be tested on other high-stakes retrieval domains, such as medical image search, where a miss is costly and set size can be interpreted as a second-opinion list.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. RA-FR combines blind face restoration (DiffBIR + InterLCM), DINOv1 ViT-B embeddings with GeM/GGeM pooling, and an RCPS-style calibration rule (Hoeffding upper confidence bound) to select a global scale κ that controls retrieval miscoverage at level α with confidence 1−δ. At inference, per-query set size is K(X,κ)=⌈κ·Φ[fu(X)]⌉, so retrieved sets adapt to estimated uncertainty. On IMFDB the authors report satisfying α=0.05 with average ~10 images; SCFace is used as a held-out surveillance/OOD test. The central claim is that RA-FR replaces fixed top-k by adaptively sized sets with a finite-sample risk guarantee.

Significance. The statistical tool is not new—it is the RCPS framework of Bates et al., adapted to retrieval as in RCIR. The paper's contribution is the engineering integration: restoration to reduce aleatoric uncertainty, DINOv1+GeM/GGeM features, and an empirical study of the resulting risk–efficiency trade-off. If the in-distribution calibration protocol is sound, the IMFDB result is a plausible and useful demonstration. However, the paper's headline deployment/OOD guarantee is not established by the experiments as reported; the SCFace protocol is underspecified and the formal guarantee is only for exchangeable calibration data. The paper is therefore promising but needs a major revision to make the claims match the evidence.

major comments (3)
  1. [§4.1, §4.4, Fig. 7/9] The SCFace evaluation is presented as a held-out/OOD test of surveillance robustness, but the manuscript never states where the κ used for SCFace was calibrated. If SCFace queries are not exchangeable with the calibration set, the guarantee of §3.1 does not apply; if κ was calibrated on an SCFace split, the results are not a domain-shift test. Please state the calibration protocol explicitly and either provide a shift-robust calibration method or restrict the guarantee claim to the matched-distribution case. As written, the abstract's 'guarantees ground truth inclusion' is not supported for SCFace, and §5's deferral of an OOD detector confirms this is a known limitation.
  2. [§3.1, Eqs. (5)–(7), Algorithm 1] The text derives a pointwise Hoeffding UCB and then selects κ̂ by searching a grid. A pointwise bound does not automatically control the data-dependent selected κ̂; the guarantee 'ρ(κ̂)≤α with probability 1−δ' requires a uniform or selection-adjusted argument. If the authors intend to rely on Bates et al.'s RCPS theorem, they should state it and verify its conditions (monotonicity, grid, δ definition); otherwise a union bound over the grid is needed. This is load-bearing because the central claim is the finite-sample guarantee.
  3. [§4.4, Fig. 7] The paper says the empirical risk lies in the green region, but Algorithm 1 selects κ̂ with UCB(κ̂)≤α. If the risk in Fig. 7 is computed on the same calibration set, the green region is guaranteed by construction and carries no evidential weight. If it is a held-out test set, the protocol (number of calibration samples, grid size, δ, repeated splits) must be reported to assess the claim. Please clarify which curve is plotted and add standard errors over repeated calibration/test splits.
minor comments (6)
  1. [§3.1, Eq. (4)] The uncertainty score Φ[fu(X)] should be defined and bounded away from zero; if zero, K=0 gives an empty set and certain miscoverage.
  2. [§4.3, Fig. 6] There is a typo 'thr risk control calibration process'; also specify what the shaded regions denote and whether the ECE is computed on the calibration or test split.
  3. [§4.1] The IMFDB paragraph contains a duplicated phrase 'and significant temporal changes'; please remove.
  4. [Algorithm 1] The grid K and its resolution are not defined. If a union bound is added, the relation between M (number of candidates), n, and δ must be given.
  5. [§4.5, Fig. 9] The text gives approximate values ('approximately 10 or fewer', 'roughly 70% larger') but no exact numbers or error bars; please report quantitative results in a table or in the text.
  6. [References] The RCIR reference [4] is cited as 'Proc. AAAI (2025)'; please provide full publication details or a DOI if available.

Circularity Check

0 steps flagged

No significant circularity: the formal risk guarantee is imported from independent conformal-risk-control results and the empirical validations are measurements, not self-referential fits.

full rationale

The central derivation chain is not circular. The claimed guarantee ρ(κ) ≤ α with probability 1−δ is not derived from the paper's own definitions or fitted quantities; it is explicitly imported from independent, peer-reviewed prior work: Bates et al. [3] and Cai et al. [4] (RCIR). Algorithm 1's selection of κ as the minimal value with R̂+(κ) ≤ α is the standard RCPS construction, and the Hoeffding upper-confidence bound in Eq. (6) is an external concentration inequality, not a result re-derived from the paper's own outputs. The paper does not rename a known empirical pattern as a new contribution: it extends RCIR with blind restoration, DINOv1 features, and GeM/GGeM pooling, while explicitly citing the existing risk-control framework. There are no self-citations to the authors' own prior work, no imported uniqueness theorem, and no ansatz smuggled in via citation. The empirical risk curves in Fig. 7 and average set sizes in Figs. 8–9 are evaluations of the calibrated procedure, not quantities defined to equal α by construction. The limitation stated in §5—that a test-time OOD detector is future work and κ is not dynamically adjusted under distributional shift—is a validity/external-generalization concern, not a circularity. It means the formal guarantee is only claimed for the matched calibration/deployment distribution, but that restriction does not make the in-distribution guarantee self-referential. Likewise, the SCFace OOD framing may be under-supported if κ was recalibrated on SCFace rather than transferred from IMFDB, but that is an experimental-design gap, not a reduction of an output to an input in the paper's equations. Overall, the derivation is self-contained and the central statistical content is honestly attributed to external sources.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central claim rests on RCPS statistical assumptions (i.i.d. calibration, monotone loss, valid bound after selection) plus domain assumptions about labels and restoration. κ is a fitted calibration parameter; the GeM exponent is learned. No invented entities are introduced.

free parameters (2)
  • κ (global controller scale) = not reported
    Chosen in Algorithm 1 as the smallest grid value whose calibrated UCB ≤ α; it directly sets every retrieval-set size via K_i=ceil(κ·u_i).
  • GeM pooling exponent p = not reported
    Declared learnable in §3.2; affects embedding quality and hence risk-efficiency, though not the formal guarantee.
axioms (5)
  • domain assumption Calibration set D_cal is i.i.d. from the deployment distribution.
    §3.1 'Let D_cal = {(X_i,Y_i)}... be an i.i.d. held-out calibration set'; the SCFace OOD evaluation (§4.4) assumes this transfers without correction, which is not established.
  • standard math The retrieval loss ℓ(κ,X,Y)=I(Y∉T_κ(X)) is monotonically non-increasing in κ.
    Needed for grid search over κ (Eq. 4-7, Algorithm 1); true because increasing κ only adds candidates to top-K.
  • ad hoc to paper Hoeffding's pointwise upper bound (Eq. 6) is valid simultaneously for the selected κ̂ after searching a grid.
    Algorithm 1 selects min over a grid using ln(1/δ)/(2n); no union-bound or grid-size correction is stated, so the claimed 1−δ confidence for the data-dependent κ̂ is not derived.
  • domain assumption Each query has a single well-defined true match Y (or the miss event is defined consistently for multiple matches).
    The risk definition and loss I(Y∉Top-K_i) treat Y as one label, while the intro defines risk over 'any of the true nearest neighbors'; the paper never reconciles the two.
  • domain assumption DiffBIR+InterLCM restoration preserves identity rather than hallucinating new identities.
    §3.3 asserts 'strictly adhering to identity constraints' but no identity-fidelity metric is reported; the whole retrieval benefit depends on it.

pith-pipeline@v1.3.0-alltime-deepseek · 9366 in / 16990 out tokens · 172506 ms · 2026-08-02T07:50:02.157786+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of A Step Forward Towards Trustworthy Risk-Aware Facial Retrieval (RA-FR)." pith.science (2026). https://pith.science/paper/4TMDJI7B

@misc{pith2026260716279,
  author       = {Pith},
  title        = {Pith review of: A Step Forward Towards Trustworthy Risk-Aware Facial Retrieval (RA-FR)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TMDJI7B}},
  note         = {Machine review of arXiv:2607.16279}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Facial image retrieval in unconstrained surveillance environments is a high-stakes challenge where missing a subject of interest -- a single false negative -- is simply not an option. Despite near-perfect performance on curated benchmarks, current recognition systems falter under real-world domain shifts such as low resolution, motion blur, and uncontrolled illumination (e.g., SCFace). Addressing this reliability gap, we propose Risk-Aware Facial Retrieval (RA-FR), a framework that moves beyond fixed Top-$k$ retrieval to adaptive set generation, guaranteeing ground truth inclusion within a user-specified risk level ($\alpha$) and confidence level ($1 - \delta$). Our approach integrates three core contributions: (1) reducing aleatoric uncertainty via a hybrid blind face restoration technique coupling Latent Consistency Models (InterLCM) and DiffBIR; (2) extracting discriminative, restoration-robust features via self-supervised DINOv1 ViT-B with GGeM pooling; and (3) employing conformal prediction with Hoeffding's inequality to dynamically calibrate retrieval set sizes based on query uncertainty. On the IMFDB benchmark, it consistently satisfies a 5% risk target with an average retrieval set size of approximately 10 images. By unifying domain-specific restoration, robust representation learning, and provable decision rules, RA-FR offers a pipeline that makes facial retrieval in surveillance both reliable and auditable. The code is available at: https://github.com/MuhammadEmmadSiddiqui/RA-FR.

Figures

Figures reproduced from arXiv: 2607.16279 by Muhammad Emmad Siddiqui, Muhammad Rafi.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed Risk-Aware Facial Retrieval framework. Vision Transformers (ViTs) address these structural limitations by employ￾ing global self-attention mechanisms, yielding object-centric embeddings that exhibit greater robustness to occlusion compared to CNNs [27]. In the facial domain specifically, pyramid ViT designs learn discriminative multi-scale rep￾resentations that surpass … view at source ↗
Figure 2
Figure 2. Figure 2: IMFDB samples demonstrating intra-class variability for a single subject [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SCFace surveillance samples showing cross-camera and resolution variability. Unlike synthetic benchmarks, SCFace introduces realistic optics-induced degradations and camera-dependent noise. This creates a rigorous testbed for cross-resolution matching, evaluating if our risk-controlled retrieval can gener- [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Recall@1 performance comparing MCD and BTL techniques on IMFDB. To assess robustness, we visualize t-SNE projections of embeddings from models trained on IMFDB [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE projections of embeddings on IMFDB and SCFace datasets. 4.3 Feature Uncertainty Estimator The feature uncertainty estimator fu is the engine of our risk-control mechanism. Since the global controller scale κ is fixed after calibration (Eq. 4), the dynamic sizing of retrieval sets depends entirely on the quality of fu. Accurate uncer￾tainty estimates are therefore crucial; a well-calibrated estimator … view at source ↗
Figure 6
Figure 6. Figure 6: Model calibration (top) & uncertainty densities (bottom) on IMFDB & SCFace. 4.4 Risk Control Effectiveness We evaluate the RA-FR procedure across a spectrum of pre-specified target risk levels on both IMFDB and SCFace [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Empirical vs. target risk (α) on IMFDB and SCFace across varying δ. The parameter δ governs the conservativeness of the calibration. Lower values of δ impose a stricter confidence requirement (i.e., a larger statistical safety margin). In practice, this forces the adapter to produce larger retrieval sets to guarantee risk satisfaction, thereby prioritizing safety over efficiency. This dynamic is clearly vi… view at source ↗
Figure 8
Figure 8. Figure 8: Average retrieval set size (K) vs. target risk (α) on the IMFDB dataset. Upgrading the embedding backbone reduces epistemic uncertainty and shrinks retrieval sets, but this alone is insufficient when queries exhibit large aleatoric uncertainty (e.g., heavy noise, blur, or extremely low resolution). To address input noise, we apply blind face restoration (BFR) to recover identity cues ex￾ploitable by downst… view at source ↗
Figure 9
Figure 9. Figure 9: Average K vs. risk (α) on SCFace with and without pre-processing steps. main shift. However, aggressive cropping can be detrimental for extremely low￾resolution queries. SCFace contains faces as small as 28×28 pixels; tight cropping effectively removes contextual and structural cues required by the enhancement modules. For example, at α = 0.05, BiD without cropping retrieves approx￾imately 226 images on av… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 5 linked inside Pith

  1. [1]

    city of detroit - case background and legal documents.https://www.aclu.org/cases/williams-v-city-of-detroit (2024), accessed: 2026-05-01

    American Civil Liberties Union: Williams v. city of detroit - case background and legal documents.https://www.aclu.org/cases/williams-v-city-of-detroit (2024), accessed: 2026-05-01

  2. [2]

    Foundations and Trends®in Machine Learning16(4), 494–591 (2023)

    Angelopoulos, A.N., Bates, S.: A gentle introduction to Conformal Prediction and distribution-free uncertainty quantification. Foundations and Trends®in Machine Learning16(4), 494–591 (2023)

  3. [3]

    Journal of the ACM (JACM)68(6), 1–34 (2021)

    Bates, S., Angelopoulos, A.N., Candès, E.J., Jordan, M.I.: Distribution-free, Risk- Controlling Prediction Sets. Journal of the ACM (JACM)68(6), 1–34 (2021)

  4. [4]

    In: Proc

    Cai, K., Lu, C.X., Zhao, X., Huang, W., Huang, X.: Risk controlled image retrieval. In: Proc. AAAI (2025)

  5. [5]

    In: Proc

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P.: Emerging properties in self-supervised vision transformers. In: Proc. ICCV (2021)

  6. [6]

    CBS News: Woman wrongly accused of carjacking loses lawsuit over facial recog- nition error.https://www.cbsnews.com/news/facial-recognition-lawsuit-p orcha-woodruff-detroit-carjacking/(2025), accessed: 2026-05-01

  7. [7]

    IEEE Trans- actions on Pattern Analysis and Machine Intelligence44(8), 4094–4114 (2022)

    Chen, W., Liu, Y., Wang, W., et al.: Deep image retrieval: A survey. IEEE Trans- actions on Pattern Analysis and Machine Intelligence44(8), 4094–4114 (2022)

  8. [8]

    In: Proc

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. ICLR (2021)

  9. [9]

    arXiv preprint arXiv:1912.02757 (2020)

    Fort, S., Hu, H., Lakshminarayanan, B.: Deep ensembles: A loss landscape per- spective. arXiv preprint arXiv:1912.02757 (2020)

  10. [10]

    In: Proc

    Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: Proc. ICML (2016)

  11. [11]

    In: Proc

    Geirhos,R.,Rubisch,P.,etal.:ImageNet-trainedCNNsarebiasedtowardstexture; increasing shape bias improves accuracy and robustness. In: Proc. ICLR (2019)

  12. [12]

    In: Proc

    Gragg, S.: Facial recognition at the CIA. In: Proc. SPIE (1997)

  13. [13]

    Mul- timedia Tools and Applications51(3), 863–879 (2011)

    Grgić, M., Delač, K., Grgić, S.: Scface - surveillance cameras face database. Mul- timedia Tools and Applications51(3), 863–879 (2011)

  14. [14]

    In: Proc

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR (2016)

  15. [15]

    arXiv preprint arXiv:2203.06082 (2022)

    Hedman, P., Skepetzis, V., Hernandez-Diaz, K., Bigun, J., Alonso-Fernandez, F.: Lfw-beautified: A dataset of face images with beautification and augmented reality filters. arXiv preprint arXiv:2203.06082 (2022)

  16. [16]

    arXiv preprint arXiv:2007.05906 (2020)

    Islam, K., Afzal, U.: Framework for passenger seat availability using face detection in passenger bus. arXiv preprint arXiv:2007.05906 (2020)

  17. [17]

    In: Proc

    Islam, K., Lee, S., Han, D., Moon, H.: Face recognition using shallow age-invariant data. In: Proc. IVCNZ (2021)

  18. [18]

    In: Proc

    Islam, K., Zaheer, M.Z., Mahmood, A.: Face pyramid vision transformer. In: Proc. BMVC (2022)

  19. [19]

    IEEE Transactions on Big Data7(3), 535–547 (2021)

    Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with gpus. IEEE Transactions on Big Data7(3), 535–547 (2021)

  20. [20]

    NeurIPS (2017) 16 M

    Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? In: Proc. NeurIPS (2017) 16 M. E. Siddiqui and M. Rafi

  21. [21]

    arXiv preprint arXiv:2212.04114 (2022)

    Ko, B., Kim, H.G., Heo, B., Yun, S., Chun, S., Gu, G., Kim, W.: Group generalized mean pooling for vision transformer. arXiv preprint arXiv:2212.04114 (2022)

  22. [22]

    Acta Numerica34, 579–804 (2025)

    Kuhn, D., Shafiee, S., Wiesemann, W.: Distributionally robust optimization. Acta Numerica34, 579–804 (2025)

  23. [23]

    arXiv:2502.02215 (2025)

    Li, S., et al.: Interlcm: Low-quality images as intermediate states... arXiv:2502.02215 (2025)

  24. [24]

    In: Proc

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Chen, D., et al.: Diffbir: Towards blind image restoration with generative diffusion prior. In: Proc. ECCV (2024)

  25. [25]

    In: Proc

    Lugaresi, C., Tang, J., Nash, H., McClanahan, C., et al.: Mediapipe: A framework for perceiving and processing reality. In: Proc. CVPR Workshops (2019)

  26. [26]

    In: Proc

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., et al.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Proc. ICLR (2022)

  27. [27]

    Rodrigo, M., Cuevas, C., García, N.: Comprehensive comparison between vision transformersandconvolutionalneuralnetworksforfacerecognitiontasks.Scientific Reports14(1), 21392 (2024)

  28. [28]

    In: Proc

    Setty, S., Husain, M., Beham, P., et al.: Indian movie face database: A benchmark for face recognition under wide variations. In: Proc. NCVPRIPG (2013)

  29. [29]

    Sobiecki, A., van Dijk, J., Folkertsma, H., Telea, A.: Does face restoration improve face verification? Multimedia Tools and Applications80(21), 32863–32883 (2021)

  30. [30]

    Cultural Studies35(4–5), 833–853 (2021)

    Stevens, N., Keyes, O.: Seeing infrastructure: race, facial recognition and the poli- tics of data. Cultural Studies35(4–5), 833–853 (2021)

  31. [31]

    In: Proc

    Warburg, F., Jørgensen, M., Civera, J., Hauberg, S.: Bayesian triplet loss: Uncer- tainty quantification in image retrieval. In: Proc. ICCV (2021)

  32. [32]

    In: Proc

    Yue, Z., Wang, J., Loy, C.C.: DifFace: Blind face restoration with diffused error contraction. In: Proc. NeurIPS. vol. 35, pp. 10533–10545 (2022)

  33. [33]

    IEEE Signal Processing Letters23(10), 1499–1503 (2016)

    Zhang, K., et al.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters23(10), 1499–1503 (2016)

  34. [34]

    In: Proc

    Zhou, S., Chan, K.C.K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup tables. In: Proc. NeurIPS. vol. 35, pp. 30599–30611 (2022)

  35. [35]

    In: Proc

    Zhu, F., Zhu, J., Chu, W., Zhang, X., Ji, X., Wang, C., Tai, Y.: Blind face restora- tion via integrating face shape and generative priors. In: Proc. CVPR (2022)