Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Factor-Informed Uncertainty Distillation for Gaze Estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that gaze-estimation uncertainty can be made both accurate and real-time by distilling an interpretable image-quality-to-error teacher into a single-pass student uncertainty head.

desk verdict A useful, well-written method paper with a real gain on Gaze360, but the evaluation is filtered, the ablations are missing, and the student's rho mirrors the teacher's—so the headline claim needs a qualifier. read the letter →

arxiv 2607.20072 v1 pith:5DYHI6JG submitted 2026-07-22 cs.CV cs.HC

Mohammadreza Jamalifard , Yaxiong Lei , Javier Fumanal Idocin , Parastoo Azizinezhad , Tom Foulsham , Javier Andreu-Perez This is my paper
classification cs.CVcs.HC
keywords gazeestimationuncertaintyquantificationknowledgedistillationselectivepredictionimagequalityteacher-studentlearningrankinglossreal-timeinference
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

FIUD trains a gradient-boosting teacher to predict expected gaze error from hand-crafted image-quality factors—illumination, sharpness, face geometry, head pose, and eye/iris features—then distills that signal into a lightweight neural uncertainty head using MSE distillation plus a margin-based ranking loss under a two-stage curriculum. At inference only the student runs, giving a single-pass uncertainty estimate with no factor extraction. Across ETH-XGaze, Gaze360, and MPIIFaceGaze (over 300k samples), FIUD improves Spearman rank correlation and selective-prediction utility (AUSE) compared to deep ensembles, MC dropout, and heteroscedastic regression, with the largest gains on the unconstrained Gaze360 dataset. If correct, this gives real-time gaze systems an abstention signal that matches or exceeds sampling-based methods at a fraction of the compute.

What carries the argument

The factor-based teacher is the central object: a gradient-boosting regressor trained on interpretable image-quality factors to predict expected gaze error, providing a structured 'quality-to-error' supervision signal. The student distills this signal through two complementary losses—L_distill (mean squared error between student and teacher uncertainties) and L_rank (a margin-based pairwise ranking loss that uses the teacher's ordering)—with curriculum learning that first warms up an error-prediction head, then transfers teacher knowledge. This decouples the cost of interpretability from inference: factors are computed only during training, while the deployed student runs at single-pass late

What would settle it

Measure FIUD's Spearman rank correlation and AUSE on a test set that includes frames where landmark detection fails entirely (e.g., complete occlusion or extreme blur). If the advantage over heteroscedastic NLL shrinks or reverses on those frames, the claim that FIUD improves reliability in unconstrained settings is not established.

Watch

Extended reading notes

Core claim

FIUD establishes that explicitly modeling the relationship between image quality and gaze error creates a stronger inductive bias for uncertainty estimation than purely pixel-derived methods. A gradient-boosting teacher maps interpretable quality factors to expected angular error, and a ResNet-18 student with an auxiliary error-prediction head and a 3-layer MLP uncertainty head is trained to match the teacher's values and ordering. At test time the student alone produces uncertainty in one forward pass, with no quality-factor extraction. On the evaluated subsets, FIUD achieves average Spearman correlation of 0.27 versus 0.14 for heteroscedastic NLL, and its AUSE on Gaze360 drops from 9.09 to

Load-bearing premise

The evaluation runs only on frames where landmark detection succeeds, so the demonstrated gains may not hold on the very frames—complete occlusions or detector failures—where a gaze system most needs to abstain.

Editorial extensions

If this is right

  • Single-pass uncertainty can replace sampling-based methods (ensembles, MC dropout) for gaze rejection tasks, cutting inference cost from 3.2x–16x to 1x while improving ranking quality.
  • Selective prediction improves most in unconstrained settings, meaning FIUD is most useful exactly where gaze systems currently fail and need to abstain.
  • Quality factors serve as supervisory signals rather than inputs, so the deployed model needs no extra feature-extraction pipeline at runtime.
  • The teacher-student pattern is a generalizable blueprint for other perception tasks—such as head pose or emotion recognition—whose failure modes can be named and measured.
  • The ranking loss directly optimizes the accept/reject ordering that interactive systems use, aligning training with deployment-time behavior.

Reading between the lines

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

  • Editorial inference: the same recipe could likely transfer to any perception task with nameable failure modes, provided a proxy error signal exists; the ranking loss might even work with a weaker proxy than ground-truth error, such as agreement between two views or a pretrained error predictor.
  • Editorial inference: the large Gaze360 gap (ρ ≈ 0.50 vs 0.32 for NLL) suggests that explicit quality supervision—not extra capacity—is what lets a single-pass head capture environmental failures; a testable follow-up is replacing hand-crafted factors with a learned no-reference quality embedding.
  • Editorial inference: the evaluation protocol filters to frames where landmark detection succeeds, so the demonstrated gains may not extend to total occlusions or detector failures; a fair deployment test would measure abstention behavior on raw, unfiltered webcam streams including such failures.
  • Editorial inference: because the teacher provides a ranking signal, the distillation could be run without labeled gaze error, potentially allowing FIUD-style uncertainty in settings where only weak or self-supervised error proxies exist.
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 Factor-Informed Uncertainty Distillation (FIUD), a two-stage teacher–student method for gaze-estimation uncertainty. A gradient-boosting teacher predicts expected angular gaze error from hand-crafted image/eye quality factors (illumination, sharpness, pose, eye features); a ResNet-18 student is trained with a composite loss combining distillation to the teacher's uncertainty, ranking supervision, and an auxiliary error-prediction head supervised by ground-truth angular error. At inference only the student is used, giving single-pass uncertainty at deterministic latency. Experiments on ETH-XGaze, Gaze360, and MPIIFaceGaze report Spearman rank correlation and AUSE, claiming improvements over deep ensembles, MC dropout, and heteroscedastic NLL, with the largest gains on Gaze360.

Significance. If the central empirical claim holds, FIUD would offer a practical single-pass uncertainty head that ranks gaze errors better than sampling-based methods while remaining cheap, a useful result for real-time HCI and safety-critical gaze applications. The paper ships a clear framework, reports efficiency numbers, and provides an interpretability story through the factor-to-error teacher. However, the evidence as presented is incomplete: the headline claim covers 'unconstrained settings' but the evaluation is restricted to the MediaPipe-success subset, and the design includes a direct ground-truth-error supervised auxiliary head, making the attribution of gains to the factor-based teacher unestablished without ablations. The work is a plausible contribution but needs substantial strengthening before the claims are supported.

major comments (3)
  1. [Abstract and §4.1] The abstract claims improvements in 'unconstrained settings' over '>300k samples,' but §4.1 states that all methods are trained and evaluated on the subset where MediaPipe landmark detection succeeds, explicitly excluding total detection failures (e.g., complete occlusion). These excluded frames are precisely where a deployment-oriented gaze system must abstain. The paper acknowledges this in §7, yet the abstract and introduction frame the result as covering unconstrained reliability. This is a load-bearing scope mismatch: either rephrase the claims to 'MediaPipe-detectable inputs' or augment the evaluation to include failure cases (e.g., by assigning a worst-case uncertainty to detection-failure frames) and report full-stream selective prediction.
  2. [§3.5–3.6, Eqs. (8)–(9)] The training objective includes an auxiliary error-prediction head that is directly supervised by ground-truth angular error in Stage 1 (Eq. 8) and retained in Stage 2 (Eq. 9). Because no ablation isolates the contribution of the factor-based teacher (distillation and ranking) from this direct error supervision, the paper's causal claim that interpretable quality factors drive the improvement is not established. A reader cannot tell whether the gains come from the teacher or simply from exposing the backbone to the true error signal. Please add ablations: (a) Stage 1 only, (b) Stage 2 without teacher (only auxiliary error), (c) distillation-only, (d) ranking-only, and (e) full FIUD. Without these, the central attribution is unsupported.
  3. [§3.3 and §5.1, Table 3] The teacher's held-out Spearman values are reported as 0.1434 (ETH-XGaze), 0.4947 (Gaze360), 0.1705 (MPIIFaceGaze). The FIUD student achieves 0.1319, 0.4980, and 0.1682, respectively. Thus the student does not improve over the factor-only teacher on two datasets and is worse on ETH-XGaze. This undercuts the §3.3 claim that the student 'can combine this distilled structure with appearance cues unavailable to the factor-only teacher, yielding uncertainty estimates that better align with observed failures.' The student appears to be a faithful, factor-free distillation of the teacher, which is still useful, but the presentation should be corrected and the role of appearance cues should be tempered.
minor comments (5)
  1. [§4.4] The statement that metrics were 'effectively identical across seeds' with three different random seeds (42/123/456) is surprising for deep networks. Please provide per-seed values or a precise explanation (e.g., deterministic evaluation with fixed seeds). As written, it undermines confidence in the empirical results.
  2. [§5.1 tables and captions] Table 3 caption should explicitly state that Spearman is computed on the MediaPipe-success test subset. Similarly, Table 4 should define AUSE units and note lower-is-better. Also, 'A verage' in Table 3 should be 'Average'.
  3. [§3.7] Typo: 'coarsequality-to-error' should be 'coarse quality-to-error'. Also, Eq. (1) typesetting of the NLL formula is hard to read; please use proper fractions.
  4. [§5.1 and §2.3] The paper cites Deep Deterministic Uncertainty [18] in Related Work but does not include it as a baseline. Since [18] is a relevant single-pass deterministic uncertainty method, adding it to Table 3 would strengthen the comparison. If omitted for scope reasons, state why.
  5. [§5.4 / Fig. 2] The reliability diagrams show mean error per uncertainty quartile but no error bars or statistical testing. Given the paper's emphasis on ranking, reporting rank-correlation confidence intervals (e.g., bootstrap) would help assess whether the differences are significant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FIUD's held-out teacher-student evaluation is self-contained.

full rationale

FIUD's derivation chain is self-contained. The teacher (Eq. 2) is trained only on training-split quality factors and angular-error labels, teacher outputs are used only for training/validation samples, and all reported test metrics come from the student on the held-out test split (Sec. 3.3, 'Leakage-free evaluation'). The student uncertainty head is trained by Eqs. 6-7 to match and rank teacher outputs; nothing in the equations sets sigma_student equal to the held-out error or to the reported Spearman/AUSE, so no prediction reduces to a fitted input by construction. Stage 1 (Eq. 8) is an auxiliary error-prediction warm-up on the backbone using the same labels as standard supervised gaze training and is not the uncertainty head used at inference. Self-citations ([15], [32], [46]-[53]) are contextual or future-work and do not supply a load-bearing uniqueness or ansatz. The acknowledged MediaPipe-success filtering (Sec. 4.1, Sec. 7) is a real scope limitation for the 'unconstrained' claim, and the unablated auxiliary loss (Sec. 3.7) weakens attribution of the gains, but both are correctness/over-claim concerns rather than circularity. The evaluation against external benchmarks is therefore not forced by definition.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or conserved quantities. The main burden is a set of domain assumptions about the predictive power of quality factors, the validity of the filtered evaluation subset, and the non-confounded role of the auxiliary error head. The free parameters are standard hyperparameters, though several are not fully reported.

free parameters (5)
  • lambda_distill = 0.5
    Weight for distillation loss, chosen based on validation performance (§3.5).
  • lambda_rank = 0.3
    Weight for ranking loss, chosen based on validation performance (§3.5).
  • margin m = 0.1
    Margin in ranking loss, tuned on validation (§3.5).
  • lambda_aux = not reported
    Weight for the auxiliary error loss in Stage 2 (Eq. 9) is not specified, so re-implementation requires tuning.
  • curriculum schedule (epochs and stage LR) = Stage 1: 20 epochs; Stage 2: 50 epochs; stage-specific LR not fully reported
    The curriculum schedule and learning-rate policy are chosen by hand and not fully specified (§3.6, §4.4).
assumptions (4)
  • domain assumption Quality factors such as illumination, sharpness, eye visibility, and symmetry are predictive of gaze estimation error.
    The entire teacher-student pipeline depends on this; teacher test Spearman values are modest (0.14–0.49), so the signal is weak in cross-subject settings (§3.3, §7).
  • domain assumption MediaPipe landmark detection success is a valid inclusion criterion for evaluating uncertainty in unconstrained settings.
    All training and evaluation are restricted to the MediaPipe-success subset, which excludes total occlusion and extreme degradation. The paper acknowledges this may bias evaluation (§4.1, §7).
  • domain assumption Ranking-oriented metrics (Spearman, AUSE) are sufficient to establish uncertainty quality; calibration is secondary.
    The paper explicitly prioritizes ranking for selective prediction (§4.3) and does not report calibration error; this is a defensible but domain-specific choice.
  • ad hoc to paper The auxiliary error-prediction head trained on ground-truth angular error does not confound the attribution of improvements to the factor-based teacher.
    Stage 1 and Stage 2 include direct MSE supervision against true error, and the paper leaves ablation to future work (§3.7). Without an ablation, the teacher's contribution to the student's final uncertainty is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Factor-Informed Uncertainty Distillation for Gaze Estimation." pith.science (2026). https://pith.science/paper/5DYHI6JG

@misc{pith2026260720072,
  author       = {Pith},
  title        = {Pith review of: Factor-Informed Uncertainty Distillation for Gaze Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DYHI6JG}},
  note         = {Machine review of arXiv:2607.20072}
}
read the original abstract

Deep gaze estimation works well in controlled capture but degrades in unconstrained settings, where systems must reject unreliable predictions. Single-pass uncertainty (e.g., heteroscedastic regression) infers uncertainty from pixels without explicit input-validity cues, while sampling based methods are often too costly for real time use. We propose Factor-Informed Uncertainty Distillation (FIUD), a teacher-student framework that aligns uncertainty with interpretable image-quality failure modes. A gradient-boosting teacher predicts expected gaze error from factors such as illumination, sharpness, eye visibility and symmetry; a neural student distills these signals via curriculum learning and ranking supervision into a lightweight single-pass uncertainty head. Across ETH-XGaze, Gaze360, and MPIIFaceGaze (>300k samples), FIUD improves uncertainty, error rank correlation and selective prediction versus deterministic and sampling-based baselines, with the largest gains in unconstrained settings.

Figures

Figures reproduced from arXiv: 2607.20072 by the authors.

Figure 1
Figure 1. FIUD framework. Training: quality factors are extracted via MediaPipe and used to train a gradient-boosting teacher. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Reliability diagrams across methods and datasets [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 3 linked inside Pith

  1. [1]

    Arantxa Villanueva and Rafael Cabeza. 2008. A novel gaze estimation system with one calibration point.IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)38, 4 (2008), 1123–1138

    2008

  2. [2]

    Yaxiong Lei, Songcen He, Mohamed Khamis, and Jun Ye. 2023. An End-to-End Review of Gaze Estimation and Its Interactive Applications on Handheld Mobile Devices.Comput. Surveys56, 2 (2023), 1–38

    2023

  3. [3]

    Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. 2024. Appearance-based gaze estimation with deep learning: A review and benchmark.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 12 (2024), 7509–7528

    2024

  4. [4]

    Yaxiong Lei, Xinya Gong, Shijing He, Yafei Wang, Mohamed Khamis, and Juan Ye. 2026. The People’s Gaze: Co-Designing and Refining Gaze Gestures with Users and Experts. InProceedings of the 2026 CHI conference on human factors in computing systems

    2026

  5. [5]

    Yaxiong Lei, Thomas Davies, Xinya Gong, Shijing He, and Juan Ye. 2026. Gaze- Code: Recall-Based Verification for Higher-Quality In-the-Wild Mobile Gaze Data Collection. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems

    2026

  6. [6]

    Yaxiong Lei, Hyochan Cho, Fergus Buchanan, Shijing He, Xinya Gong, Yuheng Wang, and Juan Ye. 2026. TinyGaze: Lightweight Gaze-Gesture Recognition on Commodity Mobile Devices. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems

    2026

  7. [7]

    Yaxiong Lei, Rishab Talwar, Shijing He, Xinya Gong, Yuheng Wang, Xudong Cai, Zhongliang Guo, and Juan Ye. 2026. GazeSync: A Mobile Eye-Tracking Tool for Analyzing Visual Attention on Dynamically Manipulated Content. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems

    2026

  8. [8]

    Francisco Vicente, Zehua Huang, Xuehan Xiong, Fernando De la Torre, Wende Zhang, and Dan Levi. 2015. Driver gaze tracking and eyes off the road detection system.IEEE Transactions on Intelligent Transportation Systems16, 4 (2015), 2014–2027

    2015

Show all 53 references
  1. [9]

    Anjul Patney, Marco Salvi, Joohwan Kim, Anton Kaplanyan, Chris Wyman, Nir Benty, David Luebke, and Aaron Lefohn. 2016. Towards foveated rendering for gaze-tracked virtual reality.ACM Trans. Graph.35, 6 (Nov. 2016), 1–12

    2016

  2. [10]

    Raptis, Mohamed Khamis, and Florian Alt

    Christina Katsini, Yasmeen Abdrabou, George E. Raptis, Mohamed Khamis, and Florian Alt. 2020. The Role of Eye Gaze in Security and Privacy Applications: Survey and Future HCI Research Directions. InProceedings of the 2020 CHI Conference on Human Factors in Computing Systems(Ho...

    2020

  3. [11]

    Shijing He, Yaxiong Lei, Zihan Zhang, Yuzhou Sun, Shujun Li, Chi Zhang, and Juan Ye. 2025. Identity Deepfake Threats to Biometric Authentication Systems: Public and Expert Perspectives. arXiv:2506.06825 [cs.HC] https://arxiv.org/abs/ 2506.06825

    2025 arXiv

  4. [12]

    Yaxiong Lei, Shijing He, Huining Feng, Kaixing Zhao, Mohamed Khamis, and Juan Ye. 2023. Protecting Privacy in an Era of Pervasive Camera-Based Devices: Challenges and Potential Directions. InProc. UK Mobile, Wearable and Ubiquitous Systems Research Symposium

    2023

  5. [13]

    Yaxiong Lei, Shijing He, Huining Feng, Kaixing Zhao, Mohamed Khamis, and Juan Ye. [n. d.]. Protecting Privacy in an Era of Pervasive Camera-Based Devices: Challenges and Potential Directions. ([n. d.])

  6. [14]

    Wenqi Zhong, Chen Xia, Dingwen Zhang, and Junwei Han. 2024. Uncertainty modeling for gaze estimation.IEEE Transactions on Image Processing33 (2024), 2851–2866

    2024

  7. [15]

    Yaxiong Lei, Yun Wang, Fiona Buchanan, Mingjie Zhao, Yusuke Sugano, Song- cen He, Mohamed Khamis, and Jun Ye. 2025. Quantifying the Impact of Mo- tion on 2D Gaze Estimation in Real-World Mobile Interactions. arXiv preprint. arXiv:2502.10570 [cs.CV] https://arxiv.org/abs/2502.10570

    2025 arXiv

  8. [16]

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems30

    2017

  9. [17]

    Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning. PMLR, 1050–1059

    2016

  10. [18]

    Jishnu Mukhoti, Andreas Kirsch, Joost Van Amersfoort, Philip HS Torr, and Yarin Gal. 2023. Deep deterministic uncertainty: A new simple baseline. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 24384– 24394

    2023

  11. [19]

    Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision?Advances in neural information processing systems30

    2017

  12. [20]

    Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. 2015. Appearance-based gaze estimation in the wild. InProceedings of the IEEE confer- ence on computer vision and pattern recognition. 4511–4520

    2015

  13. [21]

    Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. 2017. Mpiigaze: Real-world dataset and deep appearance-based gaze estimation.IEEE transactions on pattern analysis and machine intelligence41, 1 (2017), 162–175

    2017

  14. [22]

    Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhan- darkar, Wojciech Matusik, and Antonio Torralba. 2016. Eye tracking for everyone. InProceedings of the IEEE conference on computer vision and pattern recognition. 2176–2184

    2016

  15. [23]

    Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. 2020. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. InEuropean conference on computer vision. Springer, 365–381

    2020

  16. [24]

    Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Matusik, and Antonio Torralba. 2019. Gaze360: Physically unconstrained gaze estimation in the wild. InProceedings of the IEEE/CVF international conference on computer vision. 6912– 6921

    2019

  17. [25]

    Yihua Cheng, Hengfei Wang, Zhongqun Zhang, Yang Yue, Boeun Kim, Feng Lu, and Hyung Jin Chang. 2025. 3D Prior is All You Need: Cross-Task Few-shot 2D Gaze Estimation. InProceedings of the Computer Vision and Pattern Recognition Conference. 23891–23900

    2025

  18. [26]

    Ruicong Liu, Haofei Wang, and Feng Lu. 2025. From gaze jitter to domain adapta- tion: Generalizing gaze estimation by manipulating high-frequency components. International Journal of Computer Vision133, 3 (2025), 1290–1305

    2025

  19. [27]

    Jiawei Qin, Xucong Zhang, and Yusuke Sugano. 2025. UniGaze: Towards Universal Gaze Estimation via Large-scale Pre-Training

    2025

  20. [28]

    Qiaojie Zheng, Jiucai Zhang, and Xiaoli Zhang. 2025. Enhancing accuracy of uncertainty estimation in appearance-based gaze tracking with probabilistic Factor-Informed Uncertainty Distillation for Gaze Estimation ETRA ’26, June 01–04, 2026, Marrakesh, Morocco evaluation and cal...

    2025

  21. [29]

    Xin Cai, Jiabei Zeng, Shiguang Shan, and Xilin Chen. 2023. Source-free adap- tive gaze estimation by uncertainty reduction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22035–22045

    2023

  22. [30]

    Jiawei Qin, Takuru Shimoyama, Xucong Zhang, and Yusuke Sugano. 2025. Domain-adaptive full-face gaze estimation via novel-view-synthesis and fea- ture disentanglement.IEEE Access(2025)

    2025

  23. [31]

    Yafei Wang, Runze Yan, Yaxiong Lei, and Xianping Fu. 2025. PTGaze: Cross- Domain Gaze Estimation via Proxy Tuning. InProceedings of the 2025 Symposium on Eye Tracking Research and Applications. 1–2

    2025

  24. [32]

    Yaxiong Lei, Mingyue Zhao, Yuheng Wang, Shijing He, Yusuke Sugano, Yafei Wang, Kaixing Zhao, Mohamed Khamis, and Juan Ye. 2025. MAC-Gaze: Motion- Aware Continual Calibration for Mobile Gaze Tracking

    2025

  25. [33]

    Yusuke Sugano, Yasuyuki Matsushita, Yoichi Sato, and Hideki Koike. 2008. An incremental learning method for unconstrained gaze estimation. InEuropean conference on computer vision. Springer, 656–667

    2008

  26. [34]

    Daosong Hu, Mingyue Cui, and Kai Huang. 2025. UnMoDE: Uncertainty Modeling for Driver Gaze Estimation via Feature Disentanglement.IEEE Transactions on Intelligent Transportation Systems(2025)

    2025

  27. [35]

    Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. 2019. Ensemble distribu- tion distillation.arXiv preprint arXiv:1905.00076

    2019 arXiv

  28. [36]

    Javier Fumanal-Idocin and Javier Andreu-Perez. 2024. Ex-Fuzzy: A library for symbolic explainable AI through fuzzy logic programming.Neurocomputing599 (2024), 128048. doi:10.1016/j.neucom.2024.128048

    2024

  29. [37]

    Patrick Grother. 2008. Biometrics standards. InHandbook of Biometrics. Springer, 509–527

    2008

  30. [38]

    Emna Fourati, Wael Elloumi, and Aladine Chetouani. 2020. Anti-spoofing in face recognition-based biometric authentication using image quality assessment. Multimedia Tools and Applications79, 1 (2020), 865–889

    2020

  31. [39]

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. 2012. No- reference image quality assessment in the spatial domain.IEEE Transactions on image processing21, 12 (2012), 4695–4708

    2012

  32. [40]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. 2012. Making a “completely blind” image quality analyzer.IEEE Signal processing letters20, 3 (2012), 209–212

    2012

  33. [41]

    Mohd Faizan Ansari, Pawel Kasprowski, and Peter Peer. 2023. Person-specific gaze estimation from low-quality webcam images.Sensors23, 8 (2023), 4138

    2023

  34. [42]

    Yihua Cheng and Feng Lu. 2022. Gaze estimation using transformer. In2022 26th International Conference on Pattern Recognition (ICPR). IEEE, 3341–3347

    2022

  35. [43]

    Evangelos Ververas, Polydefkis Gkagkos, Jiankang Deng, Michail Christos Doukas, Jia Guo, and Stefanos Zafeiriou. 2024. 3DGazeNet: Generalizing 3D gaze estimation with weak-supervision from synthetic views. InEuropean Conference on Computer Vision. Springer, 387–404

    2024

  36. [44]

    Yusuke Sugano, Yasuyuki Matsushita, Yoichi Sato, and Hideki Koike. 2015. Appearance-based gaze estimation with online calibration from mouse oper- ations.IEEE Transactions on Human-Machine Systems45, 6 (2015), 750–760

    2015

  37. [45]

    Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. 2018. Uncertainty estimates and multi-hypotheses networks for optical flow. InProceedings of the European Conference on Computer Vision (ECCV). 652–667

    2018

  38. [46]

    Javier Andreu and Plamen Angelov. 2010. Real-time human activity recognition from wireless sensors using evolving fuzzy systems. InInternational Conference on Fuzzy Systems. IEEE, 1–8

    2010

  39. [47]

    Javier Andreu and Plamen Angelov. 2013. An evolving machine learning method for human activity recognition systems.Journal of Ambient Intelligence and Humanized Computing4, 2 (2013), 195–206

    2013

  40. [48]

    Mehrin Kiani, Javier Andreu-Perez, and Hani Hagras. 2022. A temporal type-2 fuzzy system for time-dependent explainable artificial intelligence.IEEE Transac- tions on Artificial Intelligence4, 3 (2022), 573–586

    2022

  41. [49]

    Javier Andreu-Perez, Celine Solnais, and Kumuthan Sriskandarajah. 2016. EALab (Eye Activity Lab): a MATLAB Toolbox for variable extraction, multivariate analysis and classification of eye-movement data.Neuroinformatics14, 1 (2016), 51–67

    2016

  42. [50]

    Prashant K Gupta and Javier Andreu-Perez. 2022. A gentle introduction and survey on computing with words (CWW) methodologies.Neurocomputing500 (2022), 921–937

    2022

  43. [51]

    Javier Fumanal-Idocin, Javier Andreu-Perez, Oscar Cordón, Hani Hagras, and Humberto Bustince. 2023. Artxai: Explainable artificial intelligence curates deep representation learning for artistic images using fuzzy techniques.IEEE Transactions on Fuzzy Systems32, 4 (2023), 1915–1926

    2023

  44. [52]

    Anirban Chowdhury and Javier Andreu-Perez. 2021. Clinical brain–computer interface challenge 2020 (CBCIC at WCCI2020): Overview, methods and results. IEEE Transactions on Medical Robotics and Bionics3, 3 (2021), 661–670

    2021

  45. [53]

    Laxmi Chaudhary, Nancy Girdhar, Deepak Sharma, Javier Andreu-Perez, An- toine Doucet, and Matthias Renz. 2023. A review of deep learning models for twitter sentiment analysis: Challenges and opportunities.IEEE Transactions on Computational Social Systems11, 3 (2023), 3550–3579

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.