Pith. sign in

REVIEW 4 major objections 5 minor 57 references

On the Burstiness of Faces in Set

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

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

Pith's one-line read Sets of faces used for recognition are routinely dominated by repeated or near-duplicate faces—burstiness—and the paper shows this degrades accuracy, then introduces burst-aware sampling and aggregation, including a quality-aware…

desk verdict A useful first study of burstiness in set-based face recognition, with a convincing training-side story and an evaluation-side attribution that is weakened by the QA-GMP design. read the letter →

arxiv 2506.20312 v1 pith:A7NUTDTA submitted 2025-06-25 cs.CV

Jiong Wang This is my paper · ORCID
classification cs.CV
keywords burstinessset-basedfacerecognitiongeneralizedmaxpoolingquality-awareaggregationqualityattentionburst-awaresamplingIJB-Cbenchmarkvideo
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 argues that burstiness—some faces appearing far more often than chance within a face set—pervades set-based face recognition and degrades it twice: frequent faces dominate training instances, hurting generalization, and they bias similarity comparison at evaluation. To counter this, it proposes three ways to detect bursty faces, based on Quickshift++ clustering, feature self-similarity, and generalized max-pooling (GMP) weights, and applies the detections as burst-aware sampling during training and burst-aware aggregation during evaluation. It also introduces quality-aware GMP (QA-GMP), which combines GMP's frequency signal with attention-based quality scores so that rare-but-low-quality faces are not over-weighted. Extensive experiments on IJB-B, IJB-C, and YTF demonstrate that suppressing burstiness improves recognition, with QA-GMP aggregation yielding new state-of-the-art results.

What carries the argument

The load-bearing object is the generalized max-pooling (GMP) weight vector, defined by solving $\alpha = (K + \lambda I_n)^{-1} \mathbf{1}_n$, where $K = XX^T$ is the Gram matrix of the set's face features and $\lambda$ is a regularizer. GMP equalizes the similarity between each element and the pooled representation, which in practice gives small weights to frequent, bursty faces and large weights to rare ones; the paper uses these weights both as a burstiness detector and, after an exponential transform, as sampling probabilities. QA-GMP changes the GMP objective to $\mathbf{X}^T F = \mathbf{1}_n + \lambda_4 \alpha^{\mathrm{att}}$, giving $\alpha_{\mathrm{qa}} = (K+\lambda I_n)^{-1}(\mathbf{1}_n + \lambda_4 \alpha^{\mathrm{att}})$, so the weights inherit the quality-awareness of the attention scores. Quickshift++ and self-similarity ($S = K\mathbf{1}_n/n$) serve as alternative detectors used mainly in training and on YTF.

What would settle it

On a face set with known ground-truth frequencies—for example, by inserting controlled numbers of duplicate or near-duplicate faces into a template—compute GMP weights and check whether the most frequent faces consistently receive the lowest weights; if the rank correlation between GMP weight and actual frequency is absent or reversed, the claim that GMP weights encode burstiness collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that burstiness exists widely in face sets and that the weights produced by generalized max-pooling are a usable frequency signal: faces that are frequent within a set receive lower GMP weights, and using those weights to sample less often in training and to contribute less in aggregation consistently improves recognition. The paper further discovers that plain GMP assigns high weights to low-quality faces, which it interprets as mistaking unusual features for rare ones; quality-aware GMP (QA-GMP) corrects this by regularizing the GMP objective with attention scores, so the final weights reflect both frequency and quality. With these two mechanisms applied at training and evaluation time, the paper reports the best published numbers on IJB-B, IJB-C, and YTF under its setup.

Load-bearing premise

The load-bearing premise is that GMP weights, computed from a set's own feature similarities, correctly indicate which faces are frequent (bursty) in that set, so down-weighting low-weight faces improves generalization; the paper supports this only with qualitative examples and downstream accuracy, never with ground-truth frequency labels.

Editorial extensions

If this is right

  • Burst-aware training sampling improves 1:1 verification TAR at FAR 1e-6 and 1e-5 on IJB-B and IJB-C over vanilla fine-tuning with both VGGFace2 and ArcFace backbones.
  • QA-GMP aggregation consistently beats vanilla sum-aggregation and plain GMP on IJB-C, and it improves over three different existing quality scores (FaceQNet, NAN, MCN), showing it plugs into existing quality estimators.
  • The proposed pipeline (GMP-based training plus QA-GMP evaluation) reports state-of-the-art results on IJB-B (49.04 TAR at FAR 1e-6) and IJB-C (90.87 TAR at FAR 1e-6) with the ArcFace backbone, and on YTF (97.04%).
  • On the new IJBC-BS bursty subset, the performance drop versus full IJB-C is smaller for GMP-trained and QA-GMP-evaluated models than for vanilla training, so the methods target the sets where burstiness hurts most.
  • Burst-aware sampling functions as an unsupervised hard-example miner: it up-weights infrequent faces that tend to have higher losses, without needing classifier scores or extra training overhead.

Reading between the lines

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

  • The author leaves implicit that the same recipe—detect intra-set frequency with GMP weights, then down-weight during aggregation—could transfer to other set-based tasks such as video retrieval or point-cloud recognition, where redundant frames or views act as bursts.
  • A natural next test would be to benchmark QA-GMP on sets constructed with controlled duplicate counts, separating genuine re-occurrence from low-quality outliers; the paper only evaluates on natural benchmarks plus the bursty subset.
  • Because GMP weights are computed from the set's own Gram matrix, they are dataset-agnostic and could serve as a plug-in frequency prior in any embedding space, not just face features.
  • If burstiness is as widespread as claimed, standard sum-aggregation baselines may be systematically under-reporting what their encoders can do; re-evaluating existing SFR models with burst-aware aggregation might update several published baselines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies burstiness in set-based face recognition (SFR). It argues that frequent or repetitive faces within a face set dominate training and evaluation, harming generalization and similarity computation. It proposes three detectors (Quickshift++ clustering, feature self-similarity, and generalized max-pooling weights), uses them for burst-aware instance sampling during fine-tuning and burst-aware aggregation at test time, and adds quality-aware GMP (QA-GMP) that regularizes the GMP objective with face-quality attention. Experiments on IJB-B, IJB-C, YTF, and a newly constructed IJBC-BS subset report consistent gains over vanilla fine-tuning and improved SOTA numbers on IJB-B and IJB-C with the ArcFace backbone.

Significance. If validated, the burstiness view provides a fresh explanation for failures in unconstrained SFR and yields simple, plug-in improvements. The paper is strong in breadth: two backbones, several baselines including P-norm and democratic aggregation, a new bursty-set benchmark protocol, and qualitative illustrations of detected bursty groups. It is less strong in establishing that the proposed proxies measure frequency rather than quality or difficulty, and in isolating the frequency component of QA-GMP. These issues are fixable with additional experiments, so the result is potentially significant but not yet fully established.

major comments (4)
  1. [§3.2, Eqs. (3)–(5) and §3.4, Eq. (12)] The transposition is inconsistent throughout the GMP derivation. With the convention X ∈ R^{n×d} and F ∈ R^{d×1}, Eq. (3) should read f_i F = 1 and Eq. (4) should read X F = 1_n, not X F^T = 1_n; Eq. (5) should use XF − 1_n. Likewise Eq. (12) should be XF = 1_n + λ4 α_att, not X^T F = ...; otherwise the left side has dimension d×1 while the right side is n×1. The solution in Eqs. (7) and (13) shows the intended linear system Kα = 1_n and Kα = 1_n + λ4 α_att, so the error is typographical but must be corrected for the main derivation to be checkable.
  2. [§3.2 and §3.3] The paper never validates that the GMP and self-similarity weights actually track frequency in a set. The only support is qualitative (Figs. 1, 3, 4) and indirect (loss correlations and final accuracy). Since the central claim is that suppressing burstiness—not hard-example mining or quality weighting—drives gains, I ask for a direct validation: e.g., construct synthetic face sets with known burst frequencies by duplicating images or using attribute/cluster labels, and report correlation or precision-recall of each detector against ground-truth frequency. Without this, the interpretation of GMP weights as 'burst degrees' is an assumption.
  3. [§3.4, Table 2, Table 4] The evaluation-stage ablation does not separate the frequency term from the quality term. Table 2 compares GMP(T)+S-sim(E), GMP(T)+GMP(E), and GMP(T)+QA-GMP(E) against GMP(T)+sum, and Table 4 compares quality-only scorers against the same scorers with QA-GMP, but no row applies the attention scores alone to the GMP(T) model. Add a 'GMP(T) + sum with quality attention only' condition (or remove the GMP frequency term from QA-GMP) so that the incremental gain attributable to frequency suppression can be isolated. Given §3.4's own observation that GMP over-weights low-quality faces, the current numbers do not rule out quality correction as the main source of QA-GMP's gains.
  4. [§4.2, §4.3] All four hyperparameters λ, λ1, λ2, λ3, λ4 are fixed at single values without sensitivity analysis or a stated validation-selection protocol. The behavior of Q-shift sampling changes qualitatively with λ1 in Eq. (8) (λ1=1 is uniform-by-cardinality, λ1=0 is uniform-by-group), so the reader cannot tell whether the reported gains are robust or finely tuned to the evaluation benchmarks. Please include sensitivity curves over a plausible range and/or a clear rule for setting these values on a separate validation split.
minor comments (5)
  1. [Page 2, Introduction] The sentence beginning 'For the Faces in the unconstrained scenario...' is grammatically incomplete; 'For the Faces' should be removed.
  2. [§3.5] The text says detailed statistics on the IJBC-BS subset are given in the supplementary, but the arXiv submission contains no supplementary; please include the subset statistics and exact selection rule in the paper or provide the supplement.
  3. [§3.4] The statement that applying burst-aware aggregation in training 'gets worse results' is not accompanied by any table or number; either add the comparison or qualify the claim.
  4. [Table 6] ArcFace-C is the sum-aggregation re-evaluation and ArcFace is the original exhaustive-matching result; the text should state explicitly that the two rows are not directly comparable.
  5. [§4.2] The parameter settings list should define the domain of each parameter, especially how λ4 is scaled relative to the attention-score range, since QA-GMP is sensitive to the scale of α_att.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's burstiness-suppression claims are empirically tested on external benchmarks rather than being equivalent to their inputs by construction.

full rationale

The paper's central claim is that burstiness exists in SFR face sets and that suppressing it via burst-aware sampling and aggregation improves recognition. This is supported by controlled ablations (Tables 1 and 2) against the held-out IJB-B, IJB-C, YTF, and IJBC-BS protocols. The GMP weights used as a burstiness detector are an intermediate product of a least-squares problem (Eqs. 5-7), not a parameter fitted to the benchmark labels; the paper explicitly states the hypothesis that these weights 'are supposed to reflect the burst degrees of samples' (Sec. 3.2), which is a heuristic assumption tested by downstream accuracy, not a definition that makes a prediction trivially true. No load-bearing self-citation chain exists: all references to Quickshift++, GMP, quality scorers, and related work are external prior results, and the paper does not invoke any uniqueness theorem from its own author. The QA-GMP modification (Eq. 12) is a new aggregation objective; its improvement over plain GMP is an empirical finding, and the observation that quality scores alone also help is a potential confound in interpreting the ablation, but confounding is not circularity. The hyperparameters lambda1-lambda4 are fixed in Sec. 4.2 rather than fitted on the reported test sets in a way that makes the claimed predictions equivalent to the inputs. Therefore, no specific reduction of a predicted result to its own input or to a self-citation can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on heuristics that tie feature-space properties to frequency, none of which are validated against ground-truth frequency. The GMP and self-similarity assumptions are introduced with 'supposed to' language. Hyperparameters lambda1-lambda4 are hand-set with no sensitivity study, and Q-shift's core parameters are unspecified.

free parameters (6)
  • lambda (GMP regularization) = 1
    Regularization term in Equation 5, set to 1 without sensitivity analysis.
  • lambda1 (Q-shift group sampling exponent) = 0.5
    Controls group sampling weight in Equation 8, set to 0.5.
  • lambda2 (self-similarity sampling exponent) = 2
    Controls conversion of self-similarity to sampling weight in Equation 9.
  • lambda3 (GMP sampling weight exponent) = 10
    Controls conversion of GMP weights to sampling weight in Equation 10.
  • lambda4 (QA-GMP quality balance) = 5
    Balances quality and frequency in Equation 12-13.
  • Q-shift kernel bandwidth/scale = unspecified
    Required for clustering but not detailed in the paper; affects group cardinalities and hence burstiness detection.
assumptions (4)
  • domain assumption GMP weights indicate burstiness: frequent elements get lower weights
    Stated in Section 3.2 as 'supposed to reflect the burst degrees'; not validated against ground-truth frequency.
  • domain assumption Frequent features are close to the set center, infrequent features are far
    Basis of self-similarity detection, stated in Section 3.2; used to derive S-sim weights.
  • domain assumption Attention scores reflect face quality
    Relied upon for QA-GMP; attention from prior works [55,54] is assumed to be quality-aware.
  • standard math The ridge-regression solution to Equation 5 is appropriate for computing GMP weights
    The least-squares formulation with regularization is standard and mathematically sound.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Burstiness of Faces in Set." pith.science (2026). https://pith.science/paper/A7NUTDTA

@misc{pith2026250620312,
  author       = {Pith},
  title        = {Pith review of: On the Burstiness of Faces in Set},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A7NUTDTA}},
  note         = {Machine review of arXiv:2506.20312}
}
read the original abstract

Burstiness, a phenomenon observed in text and image retrieval, refers to that particular elements appear more times in a set than a statistically independent model assumes. We argue that in the context of set-based face recognition (SFR), burstiness exists widely and degrades the performance in two aspects: Firstly, the bursty faces, where faces with particular attributes %exist frequently in a face set, dominate the training instances and dominate the training face sets and lead to poor generalization ability to unconstrained scenarios. Secondly, the bursty faces %dominating the evaluation sets interfere with the similarity comparison in set verification and identification when evaluation. To detect the bursty faces in a set, we propose three strategies based on Quickshift++, feature self-similarity, and generalized max-pooling (GMP). We apply the burst detection results on training and evaluation stages to enhance the sampling ratios or contributions of the infrequent faces. When evaluation, we additionally propose the quality-aware GMP that enables awareness of the face quality and robustness to the low-quality faces for the original GMP. We give illustrations and extensive experiments on the SFR benchmarks to demonstrate that burstiness is widespread and suppressing burstiness considerably improves the recognition performance.

Figures

Figures reproduced from arXiv: 2506.20312 by the authors.

Figure 1
Figure 1. Illustration of the burstiness phenomenon in training, evaluation datasets, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline to get the set representation. We improve the vanilla pipeline [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the sampling weights and corresponding losses of samples [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of the attention, GMP, and QA-GMP scores. The scores are [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the sample losses before and after training with three strate [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 55 canonical work pages

  1. [1]

    In: Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012)

    Abaza, A., Harrison, M.A., Bourlai, T.: Quality metrics for practical face recogni- tion. In: Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012). pp. 3103–3107. IEEE (2012) 2

    work page 2012

  2. [2]

    In: European conference on machine learning

    Akbani, R., Kwek, S., Japkowicz, N.: Applying support vector machines to imbal- anced datasets. In: European conference on machine learning. pp. 39–50. Springer (2004) 4

    work page 2004

  3. [3]

    In: Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society

    Amini, A., Soleimany, A.P., Schwarting, W., Bhatia, S.N., Rus, D.: Uncovering and mitigating algorithmic bias through learned latent structure. In: Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. pp. 289–295 (2019) 4, 6

    work page 2019

  4. [4]

    Image and Vision Computing 24(6), 639–647 (2006) 3

    Arandjelovi´ c, O., Cipolla, R.: An information-theoretic approach to face recogni- tion from face motion manifolds. Image and Vision Computing 24(6), 639–647 (2006) 3

    work page 2006

  5. [5]

    In: Proceedings of the IEEE con- ference on Computer Vision and Pattern Recognition

    Arandjelovic, R., Zisserman, A.: All about vlad. In: Proceedings of the IEEE con- ference on Computer Vision and Pattern Recognition. pp. 1578–1585 (2013) 4

    work page 2013

  6. [6]

    In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018)

    Cao, Q., Shen, L., Xie, W., Parkhi, O.M., Zisserman, A.: Vggface2: A dataset for recognising faces across pose and age. In: 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018). pp. 67–74. IEEE (2018) 1, 3, 8, 9, 10, 12, 13

    work page 2018

  7. [7]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Cevikalp, H., Dordinejad, G.G.: Discriminatively learned convex models for set based face recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10123–10132 (2019) 3

    work page 2019

  8. [8]

    In: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition

    Cevikalp, H., Triggs, B.: Face recognition based on image sets. In: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. pp. 2567–2573. IEEE (2010) 3

    work page 2010

Show all 57 references
  1. [9]

    Journal of artificial intelligence research 16, 321– 357 (2002) 4

    Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: synthetic mi- nority over-sampling technique. Journal of artificial intelligence research 16, 321– 357 (2002) 4

    2002

  2. [10]

    Natural Language Engineering 1(2), 163–190 (1995) 3

    Church, K.W., Gale, W.A.: Poisson mixtures. Natural Language Engineering 1(2), 163–190 (1995) 3

    1995

  3. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4690–4699 (2019) 1, 3, 9, 10, 12, 13

    2019

  4. [12]

    IEEE Transactions on Informa- tion Forensics and Security 7(4), 1204–1213 (2012) 2

    Ferrara, M., Franco, A., Maio, D., Maltoni, D.: Face image conformance to iso/icao standards in machine readable travel documents. IEEE Transactions on Informa- tion Forensics and Security 7(4), 1204–1213 (2012) 2

    2012

  5. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gong, S., Liu, X., Jain, A.K.: Mitigating face recognition bias via group adaptive classifier. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3414–3424 (2021) 4

    2021

  6. [14]

    In: 2019 International Conference on Biometrics (ICB)

    Gong, S., Shi, Y., Kalka, N.D., Jain, A.K.: Video face recognition: Component- wise feature aggregation network (c-fan). In: 2019 International Conference on Biometrics (ICB). pp. 1–8. IEEE (2019) 12

    2019

  7. [15]

    arXiv preprint arXiv:1904.12019 (2019) 12

    Gong, S., Shi, Y., Jain, A.K.: Recurrent embedding aggregation network for video face recognition. arXiv preprint arXiv:1904.12019 (2019) 12

    2019 arXiv

  8. [16]

    In: European conference on computer vision

    Guo, Y., Zhang, L., Hu, Y., He, X., Gao, J.: Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In: European conference on computer vision. pp. 87–102. Springer (2016) 10 16 Jiong Wang*

    2016

  9. [17]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016) 9

    2016

  10. [18]

    In: Seventh IEEE International Conference on Data Mining (ICDM 2007)

    He, Q., Chang, K., Lim, E.P.: Using burstiness to improve clustering of topics in news streams. In: Seventh IEEE International Conference on Data Mining (ICDM 2007). pp. 493–498. IEEE (2007) 3

    2007

  11. [19]

    In: 2019 Interna- tional Conference on Biometrics (ICB)

    Hernandez-Ortega, J., Galbally, J., Fierrez, J., Haraksim, R., Beslay, L.: Faceqnet: Quality assessment for face recognition based on deep learning. In: 2019 Interna- tional Conference on Biometrics (ICB). pp. 1–8. IEEE (2019) 2, 11, 12

    2019

  12. [20]

    In: 2019 IEEE 10th International Conference on Biometrics Theory, Applications and Systems (BTAS)

    Howard, J.J., Sirotin, Y.B., Vemury, A.R.: The effect of broad and specific demo- graphic homogeneity on the imposter distributions and false match rates in face recognition algorithm performance. In: 2019 IEEE 10th International Conference on Biometrics Theory, Applications a...

    2019

  13. [21]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018) 9

    2018

  14. [22]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Huang, Z., Wang, R., Shan, S., Chen, X.: Projection metric learning on grassmann manifold with application to video based face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 140–149 (2015) 3

    2015

  15. [23]

    In: 2009 IEEE conference on computer vision and pattern recognition

    J´ egou, H., Douze, M., Schmid, C.: On the burstiness of visual elements. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1169–1176. IEEE (2009) 3

    2009

  16. [24]

    In: 2010 IEEE computer society conference on computer vision and pattern recognition

    J´ egou, H., Douze, M., Schmid, C., P´ erez, P.: Aggregating local descriptors into a compact image representation. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3304–3311. IEEE (2010) 3

    2010

  17. [25]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    J´ egou, H., Zisserman, A.: Triangulation embedding and democratic aggregation for image search. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3310–3317 (2014) 4, 11

    2014

  18. [26]

    In: International Conference on Machine Learning

    Jiang, H., Jang, J., Kpotufe, S.: Quickshift++: Provably good initializations for sample-based mean shift. In: International Conference on Machine Learning. pp. 2294–2303. PMLR (2018) 2, 6

    2018

  19. [27]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Koniusz, P., Zhang, H., Porikli, F.: A deeper look at power normalizations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5774–5783 (2018) 4

    2018

  20. [28]

    In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003

    Lee, K.C., Ho, J., Yang, M.H., Kriegman, D.: Video-based face recognition using probabilistic appearance manifolds. In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 1, pp. I–I. IEEE (2003) 3

    2003

  21. [29]

    In: Asian conference on computer vision

    Li, H., Hua, G., Shen, X., Lin, Z., Brandt, J.: Eigen-pep for video face recognition. In: Asian conference on computer vision. pp. 17–33. Springer (2014) 12

    2014

  22. [30]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Liu, Y., Yan, J., Ouyang, W.: Quality aware network for set to set recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5790–5799 (2017) 2, 3, 12

    2017

  23. [31]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops

    Liu, Z., Hu, H., Bai, J., Li, S., Lian, S.: Feature aggregation network for video face recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0–0 (2019) 12

    2019

  24. [32]

    In: Proceedings of the 22nd international conference on Machine learning

    Madsen, R.E., Kauchak, D., Elkan, C.: Modeling word burstiness using the dirich- let distribution. In: Proceedings of the 22nd international conference on Machine learning. pp. 545–552 (2005) 3 On the Burstiness of Faces in Set 17

    2005

  25. [33]

    In: 2018 International Conference on Biometrics (ICB)

    Maze, B., Adams, J., Duncan, J.A., Kalka, N., Miller, T., Otto, C., Jain, A.K., Niggel, W.T., Anderson, J., Cheney, J., et al.: Iarpa janus benchmark-c: Face dataset and protocol. In: 2018 International Conference on Biometrics (ICB). pp. 158–165. IEEE (2018) 2, 9, 10

    2018

  26. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 2158–2164 (2020) 4

    Morales, A., Fierrez, J., Vera-Rodriguez, R., Tolosana, R.: Sensitivenets: Learning agnostic representations with application to face images. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(6), 2158–2164 (2020) 4

    2020

  27. [35]

    arXiv preprint arXiv:1608.06048 (2016) 4

    More, A.: Survey of resampling techniques for improving classification performance in unbalanced datasets. arXiv preprint arXiv:1608.06048 (2016) 4

    2016 arXiv

  28. [36]

    IEEE transactions on pattern analysis and machine intelligence 39(9), 1797–1810 (2016) 2, 4, 6

    Murray, N., J´ egou, H., Perronnin, F., Zisserman, A.: Interferences in match kernels. IEEE transactions on pattern analysis and machine intelligence 39(9), 1797–1810 (2016) 2, 4, 6

    2016

  29. [37]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Murray, N., Perronnin, F.: Generalized max pooling. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2473–2480 (2014) 2, 4, 6

    2014

  30. [38]

    In: European conference on computer vision

    Perronnin, F., S´ anchez, J., Mensink, T.: Improving the fisher kernel for large-scale image classification. In: European conference on computer vision. pp. 143–156. Springer (2010) 4, 11

    2010

  31. [39]

    In: Proceedings of the IEEE international conference on computer vision

    Rao, Y., Lin, J., Lu, J., Zhou, J.: Learning discriminative aggregation network for video-based face recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3781–3790 (2017) 12

    2017

  32. [40]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815–823 (2015) 1, 3, 12

    2015

  33. [41]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Shi, M., Avrithis, Y., J´ egou, H.: Early burst detection for memory-efficient image retrieval. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 605–613 (2015) 3, 5

    2015

  34. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6902–6911 (2019) 12

    2019

  35. [43]

    In: Proceedings of the IEEE conference on com- puter vision and pattern recognition

    Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: Proceedings of the IEEE conference on com- puter vision and pattern recognition. pp. 761–769 (2016) 8

    2016

  36. [44]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Sun, Y., Wang, X., Tang, X.: Deeply learned face representations are sparse, selec- tive, and robust. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2892–2900 (2015) 12

    2015

  37. [45]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Taigman, Y., Yang, M., Ranzato, M., Wolf, L.: Deepface: Closing the gap to human- level performance in face verification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1701–1708 (2014) 1, 3, 12

    2014

  38. [46]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 883–890 (2013) 3

    2013

  39. [47]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5265–5274 (2018) 12

    2018

  40. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, M., Deng, W.: Mitigating bias in face recognition using skewness-aware reinforcement learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9322–9331 (2020) 4

    2020

  41. [49]

    18 Jiong Wang* In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, W., Wang, R., Huang, Z., Shan, S., Chen, X.: Discriminant analysis on riemannian manifold of gaussian distributions for face recognition with image sets. 18 Jiong Wang* In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2048–2057 (2015) 3

    2015

  42. [50]

    In: proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Whitelam, C., Taborsky, E., Blanton, A., Maze, B., Adams, J., Miller, T., Kalka, N., Jain, A.K., Duncan, J.A., Allen, K., et al.: Iarpa janus benchmark-b face dataset. In: proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 90–98 (2017) 9, 10

    2017

  43. [51]

    In: CVPR 2011

    Wolf, L., Hassner, T., Maoz, I.: Face recognition in unconstrained videos with matched background similarity. In: CVPR 2011. pp. 529–534. IEEE (2011) 9, 10

    2011

  44. [52]

    arXiv preprint arXiv:2009.00603 (2020) 2, 3, 12

    Xie, W., Byrne, J., Zisserman, A.: Inducing predictive uncertainty estimation for face recognition. arXiv preprint arXiv:2009.00603 (2020) 2, 3, 12

    2020 arXiv

  45. [53]

    In: Proceedings of the European conference on computer vision (ECCV)

    Xie, W., Shen, L., Zisserman, A.: Comparator networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 782–797 (2018) 8, 12

    2018

  46. [54]

    BMVC (2018) 2, 3, 4, 5, 11, 12

    Xie, W., Zisserman, A.: Multicolumn networks for face recognition. BMVC (2018) 2, 3, 4, 5, 11, 12

    2018

  47. [55]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Yang, J., Ren, P., Zhang, D., Chen, D., Wen, F., Li, H., Hua, G.: Neural aggregation network for video face recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4362–4371 (2017) 2, 3, 5, 11, 12

    2017

  48. [56]

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

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

    2016

  49. [57]

    In: Asian conference on computer vision

    Zhong, Y., Arandjelovi´ c, R., Zisserman, A.: Ghostvlad for set-based face recogni- tion. In: Asian conference on computer vision. pp. 35–50. Springer (2018) 3, 4, 12

    2018

Pith tools

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