Pith. sign in

REVIEW 4 major objections 6 minor 13 references

PLGSA-Transformer: Periocular Landmark-Guided Attention with Occlusion-Adaptive Cosine Thresholding for Cross-Modal Masked and Unmasked Face Recognition

T0 review · 4 major / 6 minor · reviewed 2026-07-15 · grok-4.5

Pith's one-line read Guiding face matching with periocular landmarks and an occlusion-adaptive threshold reaches 97.2% verification accuracy and perfect ROC AUC on paired masked-unmasked images.

desk verdict Solid engineering on periocular masked matching with a real attention prior; headline AUC 1.0 on 858 partly private pairs and non-reimplemented baselines is the soft spot, not the architecture. read the letter →

arxiv 2607.03581 v2 pith:IUS7WIUX submitted 2026-07-03 cs.CV cs.AI

classification cs.CVcs.AI
keywords maskedfacerecognitionperiocularlandmarkattentionhybridCNN-Transformerocclusion-adaptivethresholdcosinesimilarityEfficientNetB3cross-modalmatchingbiometricsecurity
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

When the lower face is covered, conventional face recognizers still look at the whole crop and apply one fixed similarity cutoff. This paper argues that those two design choices are the real obstacles. It builds a matcher that first forces attention onto the eyes, brows, and forehead by injecting MediaPipe landmark heatmaps through a residual gate, then lets a short Transformer layer relate those regions to one another, and finally raises or lowers the cosine decision threshold according to a jointly learned occlusion score. On an identity-disjoint partition of an 858-image paired collection assembled for the study, the system accepts 97.2% of genuine pairs, rejects every impostor pair, and places the correct identity in the top five gallery neighbors 96% of the time. The authors present this as evidence that an explicit anatomical prior plus a severity-aware threshold closes the gap between lab numbers and deployable masked-face matching.

What carries the argument

Periocular Landmark-Guided Spatial Attention (PLGSA): MediaPipe-derived Gaussian heatmaps over 28 eye-and-brow landmarks are fused with EfficientNetB3 feature maps through a learnable residual gate (initialized at zero) so the network is steered toward the anatomically guaranteed visible zone without early-training instability.

What would settle it

Re-evaluate the identical architecture on a larger, identity-disjoint paired set that contains multiple mask types and graded coverage levels per person; a sharp drop in ROC AUC or the reappearance of false accepts would falsify the claim that the reported separability generalizes.

Watch

Extended reading notes

Core claim

Encoding periocular landmark geometry directly into spatial attention, combining it with hybrid CNN-Transformer dependency modelling, and replacing a fixed cosine cutoff with an occlusion-adaptive threshold yields a cross-modal embedding that separates genuine from impostor pairs perfectly (ROC AUC 1.0000) and reaches 97.2% pair-verification accuracy on a purpose-built 858-image masked-unmasked dataset.

Load-bearing premise

The 858-image collection built for the study, with limited per-person mask diversity and only binary mask labels, is representative enough that the perfect ROC AUC and 97.2% accuracy will hold for other people and other occluders.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes PLGSA-Transformer for cross-modal masked/unmasked face matching. It combines three components: (1) Periocular Landmark-Guided Spatial Attention (PLGSA), which injects MediaPipe-derived Gaussian heatmaps over eye/brow landmarks into EfficientNetB3 features via a learnable residual gate (Eqs. 2–3); (2) a Hybrid CNN–Transformer branch that tokenizes attended 5×5 feature maps and applies a two-layer Pre-LN MHSA encoder; and (3) Occlusion-Adaptive Cosine Thresholding (OACT), τ_i = τ(1 + β·s_i) (Eq. 7), driven by a jointly trained binary occlusion head. On an identity-disjoint partition of a purpose-built 858-image paired collection (Zenodo MDMFR 60%, Kaggle CelebA-HQ masked 25%, author-captured 15%), the model reports 97.2% pair-verification accuracy, ROC AUC 1.0000 (zero false accepts on 858 impostor pairs), Rank-1 78.0%, and Rank-5 96.0%, exceeding the same-author MUFM VGG-16 baseline (95%) and two classical methods as reported in their original papers.

Significance. If the reported separability holds under stronger evaluation, the work would be a useful, interpretable contribution to periocular masked-face matching: explicit anatomical priors in spatial attention remain underused relative to purely data-driven modules such as CBAM, and the residual-gated heatmap fusion (Eqs. 2–3) is a clean, stable design. The paper is unusually candid about data scarcity, OACT’s narrow score range (0.43–0.52), and the modest 3–4% threshold adjustment, and it carefully scopes novelty (new mechanism vs. integration). Those strengths matter. At present, however, significance is bounded by the small custom paired set, non-reimplemented baselines, and an extreme AUC of 1.0000 without uncertainty estimates, so the claim of a “robust, scalable solution” is not yet established at the level the abstract asserts.

major comments (4)
  1. [§5.1, §5.7, Table 4] §5.1 / §5.7 / Table 4: The central claim (97.2% pair accuracy, ROC AUC 1.0000, zero FPs on 858 impostor pairs) rests on a single purpose-built 858-image collection with a 15% author-captured fraction and limited per-identity occluder diversity (§3.1, §5.6, §6). Perfect separability at this scale is fragile without confidence intervals, multiple random seeds, or an external public paired benchmark. Please report bootstrap CIs (or multi-seed means±std) for AUC and accuracy, and ideally evaluate on at least one additional public protocol (e.g., synthetic-masked IJB-B/C or a held-out public paired set) so the headline numbers are not sole evidence of generalization.
  2. [Table 4, §5.7] Table 4 / §5.7: Baseline accuracies for Adnan et al. (85.0%), Shnain et al. (86.61%), and Abdullah et al. MUFM (95.0%) are taken from the original publications under different datasets and protocols. That is not a controlled comparison and cannot support the claim that PLGSA-Transformer “surpasses” those methods on the same task. Reimplement at least the same-author MUFM (and preferably one classical and one modern deep baseline) on the identical 858-pair identity-disjoint splits, or reframe Table 4 as a literature situating table rather than a head-to-head ranking.
  3. [§5.6, Eq. (7)] §5.6 / Eq. (7): OACT is presented as a core contribution, yet the occlusion head collapses to a narrow band (0.43–0.52), yielding only a 3–4% threshold shift (β=0.08). The paper is candid about this, but the abstract and conclusion still list OACT as a primary reason the system is “robust.” Either (i) provide an ablation of fixed-τ vs. OACT on the same embeddings with matched operating points, or (ii) demote OACT in the contribution list until graded multi-occluder supervision is available, so the measured effect matches the claim weight.
  4. [§4.1, §3.8] §4.1 / §3.8: Identity-disjoint 75/25 CV is described, yet test-time counts (858 pairs, 858-identity gallery) equal the full dataset size because every identity is held out once. Clarify fold construction, whether source (Zenodo/Kaggle/author) is stratified, and whether MediaPipe heatmap generation or pairing metadata could leak source-specific cues that inflate genuine/impostor separation. A source-held-out or leave-one-source-out experiment would substantially strengthen the evaluation.
minor comments (6)
  1. [Abstract / §5.1] Abstract states 97.22% while §5.1 and Table 4 use 97.2%; pick one and use consistently.
  2. [Figure 9, §5.3] Figure 9: training pair accuracy ~0.50 vs validation ~0.85–0.87 is explained by per-batch thresholds, but the figure caption should state this more prominently so readers do not misread underfitting.
  3. [§3.7, Eq. (9)] Eq. (9): center-loss weight λ₃=0.003 is two orders of magnitude smaller than λ₁/λ₂; briefly justify the scale or report a short sensitivity check.
  4. [Table 1] Table 1 lists Talahua et al. 99.65% with an asterisk for mask detection only; ensure the same caveat appears wherever that number is cited in prose.
  5. [§3.2] §3.2 lists “Color Space Conversion” and optional grayscale; EfficientNetB3 expects RGB—clarify that grayscale is not used for the reported model.
  6. [Abstract / §3.6] Minor wording: “raises the matching threshold” in the abstract vs. “relaxed threshold” for high occlusion in §3.6—align the direction of adaptation with Eq. (7).

Circularity Check

1 steps flagged · score 1.0 of 10

No definitional circularity: architecture and losses are independent of the reported metrics; only ordinary same-author baseline comparison.

  1. self citation load bearing [Section 5.7 / Table 4; also Abstract and Section 1 contribution (4)]
    "the proposed model reaches 97.2% pair verification accuracy on the present 858-image dataset, compared with the 95.0% reported by the same-author VGG-16-based MUFM model (Abdullah et al., 2025), 86.61% for the Feature-based Structural Measure (Shnain et al., 2017), and 85.0% for HOG-based classification (Adnan et al., 2020)."

    The primary quantitative claim of improvement is framed against the authors' own prior MUFM result (and other papers' self-reported numbers under different datasets/protocols). This is ordinary self-citation for incremental work and is not load-bearing for the architecture itself; the embedding and OACT are independently defined. It slightly elevates circularity only because the headline 'surpassing' rests partly on non-reimplemented, non-shared-protocol baselines rather than an external fixed benchmark.

full rationale

PLGSA-Transformer is an engineering architecture paper, not a first-principles derivation. The three components (MediaPipe Gaussian heatmaps fused via residual gate γ, Hybrid CNN-Transformer tokens, OACT τ_i = τ(1+β·s_i)) are design choices with free hyperparameters (β=0.08, λ weights, σ=10) that do not force the verification accuracy or AUC by construction. The multi-task loss (contrastive + classification + occlusion CE + center) is standard and does not define the test metrics. The only self-reference is comparison to the same-author MUFM VGG-16 predecessor (Abdullah et al., 2025, 95.0%) and citation of numbers from prior papers under their own protocols (Table 4); that is ordinary incremental work, not a load-bearing uniqueness theorem or a fitted constant renamed as prediction. OACT's narrow score band (0.43–0.52) is candidly reported as a data limitation, not smuggled as a prediction. Perfect AUC 1.0000 on the purpose-built 858-image set raises generalization/correctness concerns, but those are not circularity. Score 1 for the minor same-author baseline only.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on a small set of free hyperparameters (β, σ, loss weights, learning-rate schedule), standard domain assumptions about periocular discriminability and MediaPipe reliability, and two invented architectural entities (PLGSA gate, OACT head) whose independent evidence is limited to the same small dataset. No new physical or mathematical constants are postulated.

free parameters (5)
  • OACT scaling factor β = 0.08
    Hand-chosen conservative value 0.08 that directly scales every per-probe threshold; not derived from data or theory.
  • Gaussian heatmap σ = 10 px
    Fixed at 10 pixels for all landmarks; controls spatial support of the anatomical prior.
  • Multi-task loss weights λ1, λ2, λ3 = 0.30, 0.20, 0.003
    Set to 0.30, 0.20, 0.003 without reported sensitivity analysis; control relative influence of classification, occlusion, and center losses.
  • Base cosine threshold τ (Youden) = 0.355 (cosine distance)
    Chosen once on validation ROC; all adaptive thresholds are multiples of this fitted operating point.
  • Learning rate and schedule = 5e-5, patience 15
    Initial 5e-5 with plateau halving; affects final embedding quality on the small set.
assumptions (4)
  • domain assumption The periocular region (eyes, brows, upper nose bridge) retains sufficient identity signal for reliable matching when the lower face is occluded.
    Stated throughout Sections 1–2 and used as the justification for PLGSA; supported by prior literature but not re-proven here.
  • domain assumption MediaPipe's 478-landmark detector reliably localizes the 28 selected periocular points under the mask types and poses present in the dataset; fallback to top-third crop is adequate when detection fails.
    Section 3.3; no quantitative failure-rate analysis is supplied.
  • domain assumption Cosine similarity on L2-normalized 128-d embeddings is an appropriate verification metric, and a single global base threshold can be safely rescaled by a scalar occlusion score.
    Standard in face recognition (Section 2.4) and adopted without alternative metrics.
  • ad hoc to paper Identity-disjoint 75/25 splits of the 858-image collection, with test-time augmentation (49 gallery / 9 probe views), yield unbiased estimates of open-set performance.
    Sections 3.8 and 4.1; the small absolute size and mixed public/private sources make this a paper-specific modeling choice.
invented entities (2)
  • PLGSA residual-gated landmark attention (Eqs. 2–3)
    purpose: Inject an explicit anatomical spatial prior into CNN feature maps so attention is forced onto the visible periocular zone.
    Primary claimed architectural novelty; independent evidence is limited to qualitative heatmaps and end-to-end accuracy on the same small set; no external ablation or transfer study.
  • Occlusion-Adaptive Cosine Threshold (OACT, Eq. 7)
    purpose: Replace a fixed verification threshold with a per-probe threshold scaled by a jointly trained occlusion score.
    New decision formula; on the reported data the occlusion scores collapse to a narrow band so the entity exerts only a 3–4% effect, and no graded occlusion labels exist to test it further.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PLGSA-Transformer: Periocular Landmark-Guided Attention with Occlusion-Adaptive Cosine Thresholding for Cross-Modal Masked and Unmasked Face Recognition." pith.science (2026). https://pith.science/paper/IUS7WIUX

@misc{pith2026260703581,
  author       = {Pith},
  title        = {Pith review of: PLGSA-Transformer: Periocular Landmark-Guided Attention with Occlusion-Adaptive Cosine Thresholding for Cross-Modal Masked and Unmasked Face Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUS7WIUX}},
  note         = {Machine review of arXiv:2607.03581}
}
read the original abstract

The widespread adoption of facial masks, accelerated by COVID-19 and mandated in security-sensitive settings, has exposed limitations of conventional face recognition systems. Existing approaches relying on fixed cosine thresholds, non-adaptive CNNs, and purely data-driven features fail to generalize when facial regions are occluded, creating a gap between lab performance and real-world deployability. This paper proposes PLGSA-Transformer, a cross-modal face matching framework with three contributions. First, Periocular Landmark-Guided Spatial Attention (PLGSA) uses MediaPipe landmarks to compute Gaussian heatmaps over the eye, brow, and forehead regions, fusing them with EfficientNetB3 features via a learnable residual gate to direct attention toward discriminative visible regions. Second, a Hybrid CNN-Transformer Branch reshapes feature maps into tokens processed by a two-layer Multi-Head Self-Attention encoder, enabling cross-regional dependency modelling. Third, the Occlusion-Adaptive Cosine Threshold (OACT) is a jointly trained head that raises the matching threshold in proportion to predicted occlusion severity. The model is evaluated on 858 images from Zenodo MDMFR (60%), Kaggle CelebA-HQ masked collection (25%), and author-collected images (15%), spanning both genders, ages 21-75, with varied mask types, trained via a unified loss combining contrastive verification, identity classification, and occlusion cross-entropy. PLGSA-Transformer achieves 97.22% pair verification accuracy with ROC AUC 1.0000, surpassing VGG-16-based MUFM (Abdullah et al., 2025; 95.0%), HOG classifiers (Adnan et al., 2020; 85.0%), and Feature-based Structural Measure (Shnain et al., 2017; 86.61%). These results confirm that encoding periocular geometry into attention, with Transformer modelling and occlusion-adaptive thresholds, yields a robust, scalable solution for cross-modal masked face recognition.

Figures

Figures reproduced from arXiv: 2607.03581 by the authors.

Figure 1
Figure 1. Overview of the proposed PLGSA-Transformer pipeline. A masked–unmasked image pair of the same identity is processed by a shared (Siamese) encoder: periocular ROI extraction and a MediaPipe landmark heatmap (σ = 10) feed an EfficientNet-B3 backbone, followed by the Periocular Landmark-Guided Spatial Attention module (PLGSA, N1), a CBAM block, a 1×1 token projection to 25 tokens (d = 256) with positional encoding, and… view at source ↗
Figure 1
Figure 1. Overview of the proposed PLGSA-Transformer pipeline. A masked-unmasked image pair of the same identity is processed by a shared (Siamese) encoder: periocular ROI extraction and a MediaPipe landmark heatmap (σ = 10) feed an EfficientNet-B3 backbone, followed by the Periocular Landmark-Guided Spatial Attention module (PLGSA, N1), a CBAM block, a 1×1 token projection to 25 tokens (d = 256) with positional encoding, and… view at source ↗
Figure 2
Figure 2. Dataset collection sources. The dataset was assembled from three independent sources to ensure [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figures from the paper (19 more)
Figure 2
Figure 2. Figure 2: Dataset collection sources. The dataset was assembled from three independent sources to ensure [PITH_FULL_IMAGE:figures/full_fig_p012_2.png]
Figure 3
Figure 3. Figure 3: Dataset collection and integration process. Images from all three sources were merged into a single [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]
Figure 3
Figure 3. Figure 3: Dataset collection and integration process. Images from all three sources were merged into a single [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]
Figure 4
Figure 4. Figure 4: Gender distribution of dataset images. The dataset contains 58% male and 42% female subjects, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 4
Figure 4. Figure 4: Gender distribution of dataset images. The dataset contains 58% male and 42% female subjects, [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Data pre-processing pipeline. Eight sequential stages transform raw collected images into a standardized, augmented, and normalized format suitable for model training. The eight pre-processing stages are as follows: • Folder Organization: Images were segregated into tw…
Figure 5
Figure 5. Figure 5: Data pre-processing pipeline. Eight sequential stages transform raw collected images into a standardized, augmented, and normalized format suitable for model training. The eight pre-processing stages are as follows [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: PLGSA-Transformer proposed model architecture. The Siamese structure processes masked and unmasked image pairs through a shared branch incorporating all three novelties. The L2-normalised embedding together with the classification and occlusion heads enables joint trai…
Figure 6
Figure 6. Figure 6: PLGSA-Transformer proposed model architecture. The Siamese structure processes masked and unmasked image pairs through a shared branch incorporating all three contributions. The L2-normalised embedding together with the classification and occlusion heads enables joint …
Figure 7
Figure 7. Figure 7: Representative dataset samples. Top row: unmasked periocular images; middle row: masked images [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 7
Figure 7. Figure 7: Representative dataset samples. Top row: unmasked periocular images; middle row: masked images [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: presents the complete pair verification results, evaluated over 858 same-person pairs and 858 impostor pairs. The model achieves a ROC AUC of 1.0000, indicating perfect separability between genuine and impostor pairs on the test set, with the Youden-optimal decision th…
Figure 8
Figure 8. Figure 8: Pair verification and gallery-probe evaluation over 858 same-person pairs and 858 impostor pairs. From left to right: ROC curve with the Youden-optimal threshold marked; gallery-probe distance distributions for Rank-1 correct and incorrect retrievals; occlusion score v…
Figure 9
Figure 9. Figure 9: Per-epoch training and validation pair accuracy. The validation pair accuracy rises over the opening epochs, peaks near 0.90, and then stabilizes around 0.80, while the training pair accuracy fluctuates around 0.50. The horizontal dashed line marks the test-set match r…
Figure 9
Figure 9. Figure 9: Per-epoch training and validation pair accuracy. The validation pair accuracy rises with early fluctuation to a local peak near 0.88 by around epoch 15, then oscillates broadly within an approximately 0.80- 0.90 band for the remainder of training with a mild upward dri…
Figure 10
Figure 10. Figure 10: Eight correct Rank-1 matches spanning diverse ages, expressions, poses, and accessories. Each panel shows, left to right, the gallery (unmasked) image, the masked probe with its PLGSA attention overlay, the periocular crops compared, and the decision metrics. These ex…
Figure 10
Figure 10. Figure 10: Eight correct Rank-1 matches spanning diverse ages, expressions, poses, and accessories. Each panel shows, left to right, the gallery (unmasked) image, the masked probe with its PLGSA attention overlay, the periocular crops compared, and the decision metrics. These ex…
Figure 11
Figure 11. Figure 11: Challenging cases reported as NOT MATCHED, in which the nearest retrieved gallery identity differs [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 11
Figure 11. Figure 11: Challenging cases reported as NOT MATCHED, in which the nearest retrieved gallery identi [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 2 linked inside Pith

  1. [1]

    and Shakor, M.Y

    Abdullah, D.A., Hamad, D.R., Maolood, I.Y., Beitollahi, H., Ameen, A.K., Aula, S.A., Abdulla, A.A. and Shakor, M.Y. and Muhamad, S.S. (2025) 'A novel facial recognition technique with focusing on masked faces', Ain Shams Engineering Journal, 16, article 103350. Abdel-Hamid, L. (2021) 'Retinal image quality as sessment using transfer learning: Spatial imag...

  2. [2]

    and Subhan, F

    Alzu'bi, A., Abuarqoub, A., Quiam, F. and Subhan, F. (2021) 'Masked face recognition using deep learning: A review', Electronics, 10(21), article

  3. [3]

    and Arunachalam, K

    34 Alzubi, J.A., Pokkuluri, K.S., Arunachalam, R., Shukla, S.K., Venugopal, S. and Arunachalam, K. (2025) 'A generative adversarial network-based accurate masked face recognition model using dual scale adaptive efficient attention network', Scientific Reports, 15(1), article 17594. Deng, H., Su, Y., Zhao, M., Li, J. and Deng, H. (2021) 'MFCosface: A maske...

  4. [4]

    and Zafeiriou, S

    Deng, J ., Guo, J., Xue, N. and Zafeiriou, S. (2019) 'ArcFace: Additive angular margin loss for deep face recognition', IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10), pp. 5962–5979. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit,...

  5. [5]

    (2022) 'Efficient masked face recognition method during the COVID-19 pandemic', Signal, Image and Video Processing, 16(3), pp

    Hariri, W. (2022) 'Efficient masked face recognition method during the COVID-19 pandemic', Signal, Image and Video Processing, 16(3), pp. 605–612. Hsu, G.S.J., Liu, C.H., Lin, C.A. and Su, B.H. (2022) 'Masked face recognition from synthesis to reality', IEEE Access, 10, pp. 37938–37952. Huang, Y.C., Rahardjo, D.A.B., Shiue, R.H. and Chen, H.H. (2024) 'Mas...

  6. [6]

    and Grundmann, M

    Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M.G., Lee, J., Chang, W.T., Hua, W., Georg, M. and Grundmann, M. (2019) 'MediaPipe: A framework for building perception pipelines', arXiv preprint arXiv:1906.08172. Nsaif, A.K., Ali, S.H.M., Jassim, K.N., Nseaf, A.K., Sulaiman, R., Al -Qaraghuli, A., Wah...

  7. [7]

    and Sürantha, N

    Salim, R.J. and Sürantha, N. (2023) 'Masked face recognition by zeroing the masked region without model retraining', International Journal of Innovative Computing, Information and Control , 19(4), pp. 1087–

  8. [8]

    and Sharma, S

    Sharma, S., Khan, M.A., Mir, H.M. and Sharma, S. (2025) 'A comprehensive survey on masked face recognition techniques using deep learning: motivations, research progress, and future challenges', ICT Express, in press. Shnain, N.A., Hussain, Z.M. and Lu, S.F. (2017) 'A feature-based structural measure: An image similarity measure for face recognition', App...

Show all 13 references
  1. [9]

    (2025) 'Robust face recognition under challenging conditions: a comprehensive review of deep learning methods and challenges', Applied Sciences, 15(17), article

    Sovet, A. (2025) 'Robust face recognition under challenging conditions: a comprehensive review of deep learning methods and challenges', Applied Sciences, 15(17), article

  2. [10]

    and Kallus, N

    Steck, H., Ekanadham, C. and Kallus, N. (2024) 'Is cosine -similarity of embeddings really about similarity?', in Proceedings of the ACM Web Conference 2024 (WWW '24), Singapore, pp. 887–895. Suravarapu, V.K. and Patil, H.Y. (2025) 'Performance evaluation of enhanced deep lear...

  3. [11]

    and Le, Q

    Tan, M. and Le, Q. (2019) 'EfficientNet: Rethinking model scaling for convolutional neural networks', in Proceedings of the 36th International Conference on Machine Learning (ICML 2019), Long Beach, CA, pp. 6105–6114. Ullah, N. and Javed, A. (2022) Face mask detection and mask...

  4. [12]

    and Yue, Y

    Zhang, J., An, D., Zhang, Y., Wang, X., Wang, X., Wang, Q., Pan, Z. and Yue, Y. (2025) 'A review on face mask recognition', Sensors, 25(2), article

  5. [13]

    and Deng, W

    Zhong, Y. and Deng, W. (2021) 'Face transformer for recognition', arXiv preprint arXiv:2103.14803

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.