Pith. sign in

REVIEW 3 major objections 5 minor 68 references

A Quality-Guided Mixture of Score-Fusion Experts Framework for Human Recognition

T0 review · 3 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 A quality-guided mixture of score-fusion experts improves whole-body biometric recognition across face, gait, and body modalities, beating fixed and learned fusion baselines on four benchmarks.

desk verdict Solid, practically useful score-fusion paper; empirical results are credible, but the quality-guided MoE mechanism and pseudo-label story are oversold. read the letter →

arxiv 2508.00053 v1 pith:32IWXC7T submitted 2025-07-31 cs.CV

Jie Zhu , Yiyang Su , Minchul Kim , Anil Jain , Xiaoming Liu This is my paper · ORCID
classification cs.CV
keywords whole-bodybiometricrecognitionscorefusionmixtureofexpertsqualityestimationpseudo-labeltripletlosspersonre-identificationopen-setbiometrics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a learnable, quality-guided score-fusion layer can improve whole-body biometric recognition without retraining the underlying face, gait, or body models. The authors argue that conventional fusion—fixed weighted averaging or grid-searched weights—ignores how score distributions and input quality vary across modalities, and that their framework adapts to those variations. They report gains in Rank-1, mAP, TAR@FAR, and FNIR@FPIR over baselines on CCVID, MEVID, LTCC, and BRIAR. If true, the result is a drop-in fusion module that upgrades existing multimodal identification systems by reweighting each modality's contribution according to predicted quality.

What carries the argument

The load-bearing machinery is the Quality Estimator (QE) trained with a pseudo-quality loss. For each frame, the QE extracts intermediate features from a pretrained backbone following the aggregation scheme of [25], reduces them to a mean and standard deviation representation, and outputs a quality weight via an MLP with sigmoid. The pseudo-label is defined in Eq. 1 as a linear ramp on the ranking position of the query among gallery templates: the better the true-match rank, the closer the target weight is to 1. This weight controls the router of a Mixture of Experts layer with $Z=2$ experts, where $p_1 = w_n$ and $p_2 = 1 - w_n$, so high-quality modalities dominate the fused score matrix; the score triplet loss of Eq. 5 enforces margin separation between match and non-match fused scores. Score fusion operates on raw similarity matrices, with Euclidean distances converted to similarities via $1/(1+\mathrm{Euc}(q,g))$.

What would settle it

Train QME on a benchmark where the pretrained face model's rankings are deliberately corrupted, e.g., by enriching the gallery with easy impostor templates for low-quality queries, so that ranking success no longer tracks image quality; if fusion performance degrades or the QE weights stop correlating with measurable quality attributes like blur or resolution, the pseudo-quality premise is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that treating score fusion as a quality-conditioned mixture-of-experts problem, trained with two new losses, yields consistent gains over both fixed-rule and trained-rule fusion baselines. A modality-specific Quality Estimator predicts a scalar quality weight for each input from intermediate backbone features, using a pseudo-quality label derived from how well the query ranks against gallery templates under a pretrained model. A router consumes these weights to combine outputs of multiple fusion experts, each specialized to a different input condition, and a score triplet loss directly pushes non-match scores down while keeping match scores above a margin. The authors show this improves performance on four whole-body benchmarks, with larger gains where face quality is poor, and ablations attribute the gains to the score triplet loss, the QE guidance, and the number of experts.

Load-bearing premise

The load-bearing premise is that a better ranking of a query against gallery templates under the pretrained model means the query images are genuinely higher quality; if ranking success comes from gallery bias, label noise, or model-specific quirks rather than image quality, the QE's weights will mislead the fusion on unseen data.

Editorial extensions

If this is right

  • Any deployment using fixed score averaging can swap in QME without retraining biometric backbones and obtain higher Rank-1, mAP, TAR@FAR, and lower FNIR@FPIR on whole-body recognition benchmarks.
  • QME extends to three modalities and to models with different similarity metrics, since Euclidean distances are converted to similarities before fusion.
  • The score triplet loss is directly aligned with verification and open-set search metrics and can replace the standard triplet loss in fusion training.
  • The QE can be trained on one model (e.g., AdaFace) and used to guide fusion with another model (e.g., CAL), showing cross-model generality.
  • Larger gains appear in face-restricted or low-quality regimes, suggesting the method is most valuable where facial input is unreliable.

Reading between the lines

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

  • Editorial inference: The same pseudo-quality ranking scheme could be applied to fuse any pretrained embedding models beyond biometrics, e.g., audio-visual speaker recognition or multimodal retrieval, whenever raw similarity scores are available.
  • Editorial inference: If ranking-based pseudo-labels are noisy, an oracle-quality ablation (ground-truth blur, pose, or occlusion labels) would reveal the ceiling of the QE and how much headroom remains.
  • Editorial inference: The score triplet loss might be adapted as a general fine-tuning objective for any score-based fusion system, including face-anti-spoofing score fusion or multi-camera tracking.
  • Editorial inference: On datasets where one modality dominates, the QE could be trained adversarially to prevent easy-gallery shortcuts, since the pseudo-quality label in Eq. 1 depends on gallery composition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes QME, a score-fusion layer for multimodal whole-body biometric recognition. For each pretrained backbone, a Quality Estimator (QE) predicts a modality quality weight from intermediate features, trained by regressing to a pseudo-label derived from the query's rank against a training gallery under that backbone. The predicted weights are used to combine the outputs of a small number of score-fusion experts, which are trained with a score triplet loss that suppresses non-match scores and enforces a margin for match scores. Experiments on CCVID, MEVID, LTCC, and BRIAR compare QME with fixed and learned fusion baselines, and ablations examine the contributions of the score loss, the QE, and the number of experts.

Significance. If the results hold, QME offers a modular, backbone-agnostic way to improve score fusion without retraining biometric models. The multi-dataset evaluation across face, gait, and body models, together with open-set metrics such as FNIR@FPIR, is a real strength, and the proposed pseudo-label approach avoids manual quality annotation. The ablations give some support for each component. However, the central mechanism depends on a gallery-relative quality definition, and the evidence does not yet establish that the quality weights generalize to novel galleries; the routing implementation is also under-specified. The practical impact is potentially high for surveillance and law-enforcement applications where heterogeneous pretrained models must be combined, but the current manuscript does not fully substantiate the claimed quality-guided mechanism.

major comments (3)
  1. [Sec. 3.1, Eq. (1)] The pseudo-quality label used to train the QE is the rank of the query against the training gallery under the pretrained model. This label conflates intrinsic input quality with gallery-relative ranking: the same image can receive a low label if the gallery contains a hard impostor or if the true match is poorly represented, and a modality that is weak on a dataset will be down-weighted at the dataset level rather than at the sample level. Because the QE output is the only quality signal that gates the experts, this is a load-bearing assumption for the claim that QME improves recognition through quality-guided fusion. The ablation in Sec. 4.4 removes QE but does not test the mechanism on a novel gallery or a different gallery composition. A concrete test would be to train the QE on one gallery split and evaluate on a disjoint gallery, or to compare QE weights against a gallery with different impostor distributions; without such evidence, the reported gains could come from dataset-level reweighting rather than sample-level quality adaptation.
  2. [Sec. 3.2/3.3 vs Sec. 4.1] The described architecture has a router N_r that takes the vector of modality quality weights w_n as input and outputs expert probabilities, but the implementation section sets Z=2 with p1 = w_n and p2 = 1-p1, where w_n appears to be a single scalar. It is not specified how per-modality weights are combined when N=3, and Tables 2-4 label variants by a single QE source (e.g., 'AdaFace-QE', 'CAL-QE'), suggesting only one modality's quality weight is used. This discrepancy makes the method non-reproducible and weakens the 'mixture-of-experts' claim; please clarify whether the routing is learned, how multi-modality weights are aggregated, and which QE source was used for each reported row.
  3. [Tables 2-4] The core empirical claim of state-of-the-art performance rests on small margins in several metrics, and the paper reports no uncertainty or significance tests for Rank-1, mAP, or TAR. For example, in Table 4 the Face-Included TAR gain over Weighted-sum is 0.5 points (84.5 vs 84.0) and R20 is tied with Min-max (96.0); in Table 3 the Rank-1 gain over AIM is 0.5 points (75.3 vs 74.8). Only FNIR is reported with median and standard deviation over random non-mated subsets. Without confidence intervals or paired statistical tests, it is not possible to distinguish these improvements from evaluation noise, which is especially important given the small size of some test sets (e.g., MEVID has 316 queries).
minor comments (5)
  1. [Tables 2-4] The label 'Weigthed-sum' should be 'Weighted-sum' in several places.
  2. [Abstract] The paper states 'Code is available at the Project Link' but no URL or supplementary code is provided; please include the link or a reproducibility description.
  3. [Sec. 3.3, Eq. (2)] Equation (2) uses Euc(q,g) without defining q and g at that point; the variables should be defined where the equation appears.
  4. [Sec. 4.1] The text says 'precompute gallery features for all training subjects' but it is not clear how this training gallery relates to the test gallery used for evaluation; please clarify the distinction.
  5. [Table 2(a)] The row 'Ours' appears alongside 'Ours (AdaFace-QE)' and 'Ours (CAL-QE)' without explaining what 'Ours' without a QE label denotes; please clarify the configuration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the quality estimator is trained on ranking-based pseudo-labels, but the final claim is evaluated on held-out test sets and no equation reduces the central result to its inputs.

full rationale

The paper's derivation chain is not circular. The Quality Estimator (QE) is trained with the pseudo-quality loss L_rank (Eq. 1), where the target is a monotone function of the query's rank under a frozen pretrained model. This is a form of self-distillation: the QE learns to predict how well the pretrained model ranks a given query against the training gallery. That label is a model-specific confidence, not a direct fit of the final evaluation metric. The fused score (Eq. 3) is a weighted combination of expert outputs, with weights produced by the QE, and the score triplet loss (Eq. 5) is optimized on training data. All reported results in Tables 2-4 are computed on held-out test sets (CCVID, MEVID, LTCC, BRIAR) using standard protocols or protocols from the authors' own prior work (e.g., open-set non-mated lists following [53]); following an evaluation protocol is not a fitted parameter. No equation or citation in the manuscript makes the final Rank-1, mAP, TAR@FAR, or FNIR@FPIR values depend by construction on the training objective or on a self-citation. The assumption that a better rank implies higher quality may be empirically debatable and generalization to new galleries is a valid concern, but that is a correctness/robustness issue, not circular reasoning. The central empirical claim is therefore self-contained and independently testable.

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

The central claim rests on two assumptions: rank-based pseudo-labels are a valid quality signal, and score distributions can be normalized and gated without retraining backbones. No new physical entities are introduced. The explicit free parameters are delta, m, and Z.

free parameters (3)
  • ranking threshold delta = 3 for CCVID, MEVID, LTCC; 20 for BRIAR
    Controls the steepness of pseudo-quality labels in Eq. 1. Chosen per dataset with no sensitivity analysis, so results may depend on this tuning.
  • score margin m = 3
    Margin in the score triplet loss Eq. 5. Fixed for all datasets but no ablation or sensitivity analysis is reported.
  • number of experts Z = 2
    Model capacity choice. The claim that increasing Z gradually improves performance is only supported by Z=1 versus Z=2 in Table 5.
assumptions (3)
  • domain assumption A higher true-match ranking in the pretrained model implies higher input quality.
    Introduced in Sec. 3.1 to construct pseudo-quality labels via Eq. 1; this is the core assumption behind the QE.
  • domain assumption Score distributions from heterogeneous models can be aligned by BatchNorm and learned experts without retraining backbones.
    Needed in Sec. 3.2-3.3 for the concatenated score matrix to be fused effectively; the paper provides no theoretical justification.
  • ad hoc to paper A single quality scalar is sufficient to gate the experts for all N-modality fusion settings.
    Sec. 4.1 sets p1=w_n and p2=1-p1, so only one modality's quality weight controls the two experts; this simplification is not separately motivated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Quality-Guided Mixture of Score-Fusion Experts Framework for Human Recognition." pith.science (2026). https://pith.science/paper/32IWXC7T

@misc{pith2026250800053,
  author       = {Pith},
  title        = {Pith review of: A Quality-Guided Mixture of Score-Fusion Experts Framework for Human Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32IWXC7T}},
  note         = {Machine review of arXiv:2508.00053}
}
read the original abstract

Whole-body biometric recognition is a challenging multimodal task that integrates various biometric modalities, including face, gait, and body. This integration is essential for overcoming the limitations of unimodal systems. Traditionally, whole-body recognition involves deploying different models to process multiple modalities, achieving the final outcome by score-fusion (e.g., weighted averaging of similarity matrices from each model). However, these conventional methods may overlook the variations in score distributions of individual modalities, making it challenging to improve final performance. In this work, we present \textbf{Q}uality-guided \textbf{M}ixture of score-fusion \textbf{E}xperts (QME), a novel framework designed for improving whole-body biometric recognition performance through a learnable score-fusion strategy using a Mixture of Experts (MoE). We introduce a novel pseudo-quality loss for quality estimation with a modality-specific Quality Estimator (QE), and a score triplet loss to improve the metric performance. Extensive experiments on multiple whole-body biometric datasets demonstrate the effectiveness of our proposed approach, achieving state-of-the-art results across various metrics compared to baseline methods. Our method is effective for multimodal and multi-model, addressing key challenges such as model misalignment in the similarity score domain and variability in data quality.

Figures

Figures reproduced from arXiv: 2508.00053 by the authors.

Figure 1
Figure 1. Illustration of score distribution alignment in multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. General framework for whole-body biometric recognition. An input video sequence [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The architecture of the proposed QME framework. It includes a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Score distributions of the CCVID test set. [Keys: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The distribution of AdaFace quality weights for the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 62 canonical work pages

  1. [1]

    Geff: improving any clothes-changing person ReID model using gallery enrichment with face features

    Daniel Arkushin, Bar Cohen, Shmuel Peleg, and Ohad Fried. Geff: improving any clothes-changing person ReID model using gallery enrichment with face features. InWACV, 2024. 6

    work page 2024

  2. [2]

    Learning face image quality from human assessments.IEEE Transactions on In- formation forensics and security, 13(12), 2018

    Lacey Best-Rowden and Anil K Jain. Learning face image quality from human assessments.IEEE Transactions on In- formation forensics and security, 13(12), 2018. 2

    work page 2018

  3. [3]

    Bio- metric quality: a review of fingerprint, iris, and face

    Samarth Bharadwaj, Mayank Vatsa, and Richa Singh. Bio- metric quality: a review of fingerprint, iris, and face. EURASIP journal on Image and Video Processing, 2014,

    work page 2014

  4. [4]

    Data uncertainty learning in face recognition

    Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. InCVPR,

    work page

  5. [5]

    Atm: Action tempo- rality modeling for video question answering

    Junwen Chen, Jie Zhu, and Yu Kong. Atm: Action tempo- rality modeling for video question answering. InACM MM,

  6. [6]

    An approach for full reinforcement-based biometric score fusion.IEEE Access, 2024

    Mohamed Cheniti, Zahid Akhtar, Chandranath Adak, and Kamran Siddique. An approach for full reinforcement-based biometric score fusion.IEEE Access, 2024. 2

    work page 2024

  7. [7]

    Expanding accurate person recognition to new altitudes and ranges: The briar dataset

    David Cornett, Joel Brogan, Nell Barber, Deniz Aykac, Seth Baird, Nicholas Burchfield, Carl Dukes, Andrew Duncan, Regina Ferrell, Jim Goddard, et al. Expanding accurate person recognition to new altitudes and ranges: The briar dataset. InWACV, 2023. 2

    work page 2023

  8. [8]

    Generalizable person re-identification with relevance-aware mixture of experts

    Yongxing Dai, Xiaotong Li, Jun Liu, Zekun Tong, and Ling-Yu Duan. Generalizable person re-identification with relevance-aware mixture of experts. InCVPR, 2021. 4

    work page 2021

Show all 68 references
  1. [9]

    Ca- bala—collaborative architectures based on biometric adapt- able layers and activities.PR, 45(6):2348–2362, 2012

    Maria De Marsico, Michele Nappi, and Daniel Riccio. Ca- bala—collaborative architectures based on biometric adapt- able layers and activities.PR, 45(6):2348–2362, 2012. 2

    2012

  2. [10]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InCVPR, 2019. 1

    2019

  3. [11]

    Quality assessment of image-based biometric information.EURASIP Journal on Image and video Process- ing, 2015, 2015

    Mohamad El-Abed, Christophe Charrier, and Christophe Rosenberger. Quality assessment of image-based biometric information.EURASIP Journal on Image and video Process- ing, 2015, 2015. 2

    2015

  4. [12]

    Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.Journal of Machine Learning Re- search, 23(120), 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.Journal of Machine Learning Re- search, 23(120), 2022. 4

    2022

  5. [13]

    Performance of biomet- ric quality measures.TPAMI, 29(4), 2007

    Patrick Grother and Elham Tabassi. Performance of biomet- ric quality measures.TPAMI, 29(4), 2007. 2

    2007

  6. [14]

    Face re-identification in thermal infrared spectrum based on Ther- malFaceNet neural network

    Artur Grudzien, Marcin Kowalski, and Norbert Palka. Face re-identification in thermal infrared spectrum based on Ther- malFaceNet neural network. InMIKON, 2018. 2

    2018

  7. [15]

    Clothes-changing person re-identification with rgb modality only

    Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. InCVPR, 2022. 1, 5, 6, 7

    2022

  8. [16]

    Multi- domain learning for updating face anti-spoofing models

    Xiao Guo, Yaojie Liu, Anil Jain, and Xiaoming Liu. Multi- domain learning for updating face anti-spoofing models. In ECCV, 2022. 2

    2022

  9. [17]

    Rethinking vision-language model in face forensics: Multi-modal interpretable forged face detector

    Xiao Guo, Xiufeng Song, Yue Zhang, Xiaohong Liu, and Xiaoming Liu. Rethinking vision-language model in face forensics: Multi-modal interpretable forged face detector. In CVPR, 2025. 2

    2025

  10. [18]

    Multi-modal human authentication using silhouettes, gait and rgb

    Yuxiang Guo, Cheng Peng, Chun Pong Lau, and Rama Chel- lappa. Multi-modal human authentication using silhouettes, gait and rgb. InFG, 2023. 2

    2023

  11. [19]

    Performance evaluation of score level fusion in multimodal biometric systems.PR, 43 (5), 2010

    Mingxing He, Shi-Jinn Horng, Pingzhi Fan, Ray-Shine Run, Rong-Jian Chen, Jui-Lin Lai, Muhammad Khurram Khan, and Kevin Octavius Sentosa. Performance evaluation of score level fusion in multimodal biometric systems.PR, 43 (5), 2010. 2, 5, 6, 7

    2010

  12. [20]

    Combining multiple biometric traits us- ing asymmetric aggregation operators for improved person recognition.Symmetry, 12(3):444, 2020

    Abderrahmane Herbadji, Zahid Akhtar, Kamran Siddique, Noubeil Guermat, Lahcene Ziet, Mohamed Cheniti, and Khan Muhammad. Combining multiple biometric traits us- ing asymmetric aggregation operators for improved person recognition.Symmetry, 12(3):444, 2020. 5, 6, 7

    2020

  13. [21]

    Faceqnet: Quality assessment for face recognition based on deep learning

    Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In ICB, 2019. 2

    2019

  14. [22]

    Whole-body detection, recognition and identification at altitude and range.arXiv preprint arXiv:2311.05725, 2023

    Siyuan Huang, Ram Prabhakar Kathirvel, Chun Pong Lau, and Rama Chellappa. Whole-body detection, recognition and identification at altitude and range.arXiv preprint arXiv:2311.05725, 2023. 2

    2023 arXiv

  15. [23]

    Score nor- malization in multimodal biometric systems.PR, 38(12),

    Anil Jain, Karthik Nandakumar, and Arun Ross. Score nor- malization in multimodal biometric systems.PR, 38(12),

  16. [24]

    Adaface: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. InCVPR,

  17. [25]

    Cluster and aggregate: Face recognition with large probe set

    Minchul Kim, Feng Liu, Anil K Jain, and Xiaoming Liu. Cluster and aggregate: Face recognition with large probe set. InNeurIPS, 2022. 2, 3, 6

    2022

  18. [26]

    KeyPoint Relative Position Encoding for Face Recog- nition

    Minchul Kim, Yiyang Su, Feng Liu, Anil Jain, and Xiaoming Liu. KeyPoint Relative Position Encoding for Face Recog- nition. InCVPR, 2024. 5, 7

    2024

  19. [27]

    Sapiensid: Foundation for human recognition

    Minchul Kim, Dingqiang Ye, Yiyang Su, Feng Liu, and Xi- aoming Liu. Sapiensid: Foundation for human recognition. InCVPR, 2025. 1, 2, 5, 6, 7

    2025

  20. [28]

    A new probabilistic iris quality measure for com- prehensive noise detection

    Emine Krichen, Sonia Garcia-Salicetti, and Bernadette Dorizzi. A new probabilistic iris quality measure for com- prehensive noise detection. InBTAS, 2007. 2

    2007

  21. [29]

    Gshard: Scaling giant models with conditional computation and automatic sharding.arXiv preprint arXiv:2006.16668, 2020

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding.arXiv preprint arXiv:2006.16668, 2020. 4

    2006 arXiv

  22. [30]

    Toward facial re- identification: Experiments with data from an operational surveillance camera plant

    Pei Li, Joel Brogan, and Patrick J Flynn. Toward facial re- identification: Experiments with data from an operational surveillance camera plant. InBTAS, 2016. 2

    2016

  23. [31]

    Learning face similarity for re-identification from real surveillance video: A deep metric solution

    Pei Li, Maria Loreto Prieto, Patrick J Flynn, and Domingo Mery. Learning face similarity for re-identification from real surveillance video: A deep metric solution. InIJCB, 2017. 2

    2017

  24. [32]

    An in-depth ex- ploration of person re-identification and gait recognition in cloth-changing conditions

    Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth ex- ploration of person re-identification and gait recognition in cloth-changing conditions. InCVPR, 2023. 1

    2023

  25. [33]

    Grid search, random search, genetic algorithm: a big comparison for nas

    Petro Liashchynskyi and Pavlo Liashchynskyi. Grid search, random search, genetic algorithm: a big comparison for nas. arXiv preprint arXiv:1912.06059, 2019. 2

    1912 arXiv

  26. [34]

    Farsight: A physics- driven whole-body biometric system at large distance and al- titude

    Feng Liu, Ryan Ashbaugh, Nicholas Chimitt, Najmul Has- san, Ali Hassani, Ajay Jaiswal, Minchul Kim, Zhiyuan Mao, Christopher Perry, Zhiyuan Ren, et al. Farsight: A physics- driven whole-body biometric system at large distance and al- titude. InWACV, 2024. 2, 5, 6, 7

    2024

  27. [35]

    Distilling CLIP with Dual Guidance for Learning Discrimi- native Human Body Shape Representation

    Feng Liu, Minchul Kim, Zhiyuan Ren, and Xiaoming Liu. Distilling CLIP with Dual Guidance for Learning Discrimi- native Human Body Shape Representation. InCVPR, 2024. 1, 5, 6, 7

    2024

  28. [36]

    Person recognition at altitude and range: Fusion of face, body shape and gait

    Feng Liu, Nicholas Chimitt, Lanqing Guo, Jitesh Jain, Aditya Kane, Minchul Kim, Wes Robbins, Yiyang Su, Dingqiang Ye, Xingguang Zhang, et al. Person recognition at altitude and range: Fusion of face, body shape and gait. arXiv preprint arXiv:2505.04616, 2025. 2

    2025 arXiv

  29. [37]

    Magface: A universal representation for face recognition and quality assessment

    Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. InCVPR, 2021. 2

    2021

  30. [38]

    Likelihood ratio-based biometric score fusion.TPAMI, 30 (2), 2007

    Karthik Nandakumar, Yi Chen, Sarat C Dass, and Anil Jain. Likelihood ratio-based biometric score fusion.TPAMI, 30 (2), 2007. 2

    2007

  31. [39]

    Improving face recognition with a quality-based probabilistic framework

    Necmiye Ozay, Yan Tong, Frederick W Wheeler, and Xi- aoming Liu. Improving face recognition with a quality-based probabilistic framework. InCVPRW, 2009. 2

    2009

  32. [40]

    Multi-scale speaker diarization with neural affinity score fu- sion

    Tae Jin Park, Manoj Kumar, and Shrikanth Narayanan. Multi-scale speaker diarization with neural affinity score fu- sion. InICASSP, 2021. 2, 5, 6, 7

    2021

  33. [41]

    A unified framework for bio- metric expert fusion incorporating quality measures.TPAMI, 34(1), 2011

    Norman Poh and Josef Kittler. A unified framework for bio- metric expert fusion incorporating quality measures.TPAMI, 34(1), 2011. 2

    2011

  34. [42]

    Im- proving biometric device interoperability by likelihood ratio- based quality dependent score normalization

    Norman Poh, Josef Kittler, and Thirimachos Bourlai. Im- proving biometric device interoperability by likelihood ratio- based quality dependent score normalization. InBTAS, 2007. 2

    2007

  35. [43]

    Long-term cloth-changing person re-identification

    Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 1

    2020

  36. [44]

    Implicit Discriminative Knowl- edge Learning for Visible-Infrared Person Re-Identification

    Kaijie Ren and Lei Zhang. Implicit Discriminative Knowl- edge Learning for Visible-Infrared Person Re-Identification. InCVPR, 2024. 2

    2024

  37. [45]

    Information fusion in biometrics

    Arun Ross and Anil Jain. Information fusion in biometrics. PR letters, 24(13), 2003. 2

    2003

  38. [46]

    Re-iqa: Unsupervised learning for image quality assessment in the wild

    Avinab Saha, Sandeep Mishra, and Alan C Bovik. Re-iqa: Unsupervised learning for image quality assessment in the wild. InCVPR, pages 5846–5855, 2023. 2

    2023

  39. [47]

    Facenet: A unified embedding for face recognition and clus- tering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. InCVPR, 2015. 5

    2015

  40. [48]

    Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer.arXiv preprint arXiv:1701.06538, 2017. 4

    2017 arXiv

  41. [49]

    Mesh-tensorflow: Deep learning for supercomputers

    Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, et al. Mesh-tensorflow: Deep learning for supercomputers. In NeurIPS, 2018. 4

    2018

  42. [50]

    Probabilistic face embeddings

    Yichun Shi and Anil K Jain. Probabilistic face embeddings. InICCV, 2019. 2

    2019

  43. [51]

    A comprehen- sive overview of biometric fusion.Information Fusion, 52,

    Maneet Singh, Richa Singh, and Arun Ross. A comprehen- sive overview of biometric fusion.Information Fusion, 52,

  44. [52]

    Multimodal biometrics: issues in design and testing

    Robert Snelick, Mike Indovina, James Yen, and Alan Mink. Multimodal biometrics: issues in design and testing. In ICMI, 2003. 5, 6, 7

    2003

  45. [53]

    Open-set biometrics: Beyond good closed-set models

    Yiyang Su, Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Open-set biometrics: Beyond good closed-set models. InECCV, 2024. 2, 6

    2024

  46. [54]

    Hamobe: Hierarchical and adaptive mixture of biometric ex- perts for video-based person reid

    Yiyang Su, Yunping Shi, Feng Liu, and Xiaoming Liu. Hamobe: Hierarchical and adaptive mixture of biometric ex- perts for video-based person reid. InICCV, 2025. 4

    2025

  47. [55]

    Optimized score level fusion for multi-instance finger vein recognition

    Jackson Horlick Teng, Thian Song Ong, Tee Connie, Kala- iarasi Sonai Muthu Anbananthen, and Pa Pa Min. Optimized score level fusion for multi-instance finger vein recognition. Algorithms, 2022. 2, 5, 6, 7

    2022

  48. [56]

    Ser-fiq: Unsupervised esti- mation of face image quality based on stochastic embedding robustness

    Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised esti- mation of face image quality based on stochastic embedding robustness. InCVPR, 2020. 2

    2020

  49. [57]

    Im- proving biometric identification through quality-based face and fingerprint biometric fusion

    Yan Tong, Frederick W Wheeler, and Xiaoming Liu. Im- proving biometric identification through quality-based face and fingerprint biometric fusion. InCVPRW, 2010. 2

    2010

  50. [58]

    Integrating image quality in 2ν-svm biometric match score fusion.In- ternational Journal of Neural Systems, 17(05), 2007

    Mayank Vatsa, Richa Singh, and Afzel Noore. Integrating image quality in 2ν-svm biometric match score fusion.In- ternational Journal of Neural Systems, 17(05), 2007. 2

    2007

  51. [59]

    Adaptive graph representation learn- ing for video person re-identification.IEEE Transactions on Image Processing, 29, 2020

    Yiming Wu, Omar El Farouk Bourahla, Xi Li, Fei Wu, Qi Tian, and Xue Zhou. Adaptive graph representation learn- ing for video person re-identification.IEEE Transactions on Image Processing, 29, 2020. 5, 6

    2020

  52. [60]

    Shallow-Deep Collab- orative Learning for Unsupervised Visible-Infrared Person Re-Identification

    Bin Yang, Jun Chen, and Mang Ye. Shallow-Deep Collab- orative Learning for Unsupervised Visible-Infrared Person Re-Identification. InCVPR, 2024. 2

    2024

  53. [61]

    Person re- identification by contour sketch under moderate clothing change.TPAMI, 43(6), 2019

    Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re- identification by contour sketch under moderate clothing change.TPAMI, 43(6), 2019. 1

    2019

  54. [62]

    Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification

    Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. InCVPR, 2023. 5, 7

    2023

  55. [63]

    BigGait: Learning Gait Representation You Want by Large Vision Models

    Dingqiang Ye, Chao Fan, Jingzhe Ma, Xiaoming Liu, and Shiqi Yu. BigGait: Learning Gait Representation You Want by Large Vision Models. InCVPR, 2024. 1, 5, 6, 7

    2024

  56. [64]

    Score level fusion of classifiers in off-line signature verification.Infor- mation Fusion, 32, 2016

    Mustafa Berkay Yılmaz and Berrin Yanıko ˘glu. Score level fusion of classifiers in off-line signature verification.Infor- mation Fusion, 32, 2016. 2

    2016

  57. [65]

    Common sense reasoning for deepfake de- tection

    Yue Zhang, Ben Colman, Xiao Guo, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deepfake de- tection. InECCV, 2024. 2

    2024

  58. [66]

    Gait recognition via disentangled representation learning

    Ziyuan Zhang, Luan Tran, Xi Yin, Yousef Atoum, Xiaom- ing Liu, Jian Wan, and Nanxin Wang. Gait recognition via disentangled representation learning. InCVPR, 2019. 1

    2019

  59. [67]

    Webface260m: A benchmark unveiling the power of million-scale deep face recognition

    Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. InCVPR,

  60. [68]

    Tam- ing sparsely activated transformer with stochastic experts

    Simiao Zuo, Xiaodong Liu, Jian Jiao, Young Jin Kim, Hany Hassan, Ruofei Zhang, Tuo Zhao, and Jianfeng Gao. Tam- ing sparsely activated transformer with stochastic experts. In ICLR, 2022. 4

    2022

Pith tools

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