Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Quantifying the Impact of Motion on 2D Gaze Estimation in Real-World Mobile Interactions

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read User motion degrades 2D phone gaze tracking, and most of the drop traces to head distance and device orientation.

desk verdict A useful empirical study with a new synchronized IMU+camera dataset; the qualitative motion-accuracy findings are solid, but the headline percentages and top factor rest on an unvalidated head-distance proxy and an internal numeric inconsistency, so treat the factor ranking as provisional. read the letter →

arxiv 2502.10570 v2 pith:OXU2LPTB submitted 2025-02-14 cs.HC cs.CV

classification cs.HCcs.CV
keywords MobileGazeTracking2DEstimationCalibrationDevicesIMUSensorsUserStudiesLassoRegressionHeadDistance
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, with quantitative evidence, that the usual assumption behind 2D gaze estimation on phones—that a model can implicitly handle whatever head pose and device motion a user produces—fails in realistic mobile use. In two user studies covering lying, sitting, standing, walking, and maze navigation, the authors measured how much accuracy drops once users move: dynamic conditions raise gaze-point error by up to 48.91% relative to static ones, and a one-time calibration leaves a mean error around 3.44 cm. Using Lasso regression on synchronized inertial and facial-image features, they attribute the largest shares of error to head-to-screen distance (about 36%), device orientation (about 32–38% depending on how it is grouped), and head movements (about 10%). If this attribution is right, the path to more accurate mobile gaze interaction is not merely better appearance models but explicit handling of head distance and device tilt, including adaptive or more frequent calibration.

What carries the argument

The load-bearing mechanism is an integrated measurement-and-analysis pipeline. A mobile app synchronizes IMU events (accelerometer, gyroscope, magnetometer at 50 Hz) with front-camera frames (30 Hz) while users perform tasks; head pose is derived from facial landmarks, and head-to-camera distance is estimated by a participant-specific linear regression on the ratio of physical interpupillary distance to its in-image length, i.e. $D_{\text{Head-Camera}} = \beta_0 + \beta_1 \cdot \frac{IOD_{\text{Physical}}}{IOD_{\text{Frame}}} + \epsilon$, calibrated with a ruler. Gaze error is computed as the Euclidean distance between predicted and ground-truth on-screen points, and a Lasso regression (with cross-validated regularization) is run on 552 engineered time, frequency, and cross-channel features to decompose error into factor groups. The key identity the paper relies on is that the Lasso coefficient sums assigned to head distance, head movement, and device orientation are large enough to dominate the model's total weight, making those three groups the practical levers for error reduction.

What would settle it

Run the same lying, sitting, standing, walking, and maze protocol while measuring true head-to-camera distance with a depth sensor or motion capture, then recompute the Lasso contribution of head distance; if the ground-truth distance's share falls below the share of the interpupillary-ratio proxy, the top factor is partly an artifact of the proxy. A simpler version: hold head distance fixed across all conditions and measure whether the predicted orientation-driven error still appears.

Watch

Extended reading notes

Core claim

The paper's central claim is that motion-induced changes in the spatial relationship between the user's head and the handheld device are the dominant, quantifiable cause of error in 2D appearance-based gaze estimation. The authors show this by having 10 participants perform five everyday tasks under five motion conditions while logging device IMU data, front-camera frames, and ground-truth gaze points, then training participant-specific SVR calibrators on the features of a base 2D gaze model. Their measurements show a systematic accuracy gradient from static to dynamic conditions—one-off calibration achieves its best mean error of 3.07 cm when sitting and worsens to 3.79 cm when walking in a maze—and a motion-specific calibration that trains on 30% of the same-condition test data drives error below 2 cm in static conditions but degrades by roughly 49% under dynamic ones. A Lasso regression over 552 sensor features identifies head-to-screen distance, head movements, and device-orientation deflection as the key contributors, with head distance and orientation together carrying the majority of the regression weight. The conclusion offered is that 2D gaze estimators should incorporate explicit head-pose and distance handling, or periodic and adaptive recalibration, rather than relying on an implicit learned tolerance for motion.

Load-bearing premise

The head-to-camera distance, which the Lasso ranks as the single most important error factor, is estimated from an interpupillary-distance ratio via a participant-specific ruler-calibrated regression and is never validated against true depth, so if that proxy is biased by head pose or lighting, the headline attribution of roughly 36% of error weight to head distance would be misestimated.

Editorial extensions

If this is right

  • One-time calibration is not enough for mobile use: the paper's one-off calibrator averages 3.44 cm RMSE and is worst (3.79 cm) in the most dynamic condition.
  • Calibrating to the specific task or motion state helps: task-specific calibration improves over one-off calibration by about 13.41% on average, and motion-specific calibration can reach under 2 cm in static conditions.
  • Errors under walking and maze navigation are significantly higher than under lying and sitting, with moderate to large effect sizes, so mobile gaze systems need to treat dynamic use as a distinct operating regime.
  • Head-to-screen distance and device-orientation deflection together outweigh all other factors in the Lasso decomposition, so estimators that feed these signals into the pipeline should see the largest accuracy gains.

Reading between the lines

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

  • If the paper's attribution is right, then continuously estimating face scale or depth and triggering recalibration only when distance shifts should recover most of the lost accuracy at a fraction of the cost of continuous calibration.
  • The abstract's 32.29% and the body's 38.26% figures for device orientation suggest the umbrella term 'device orientation' is doing different work in different places, namely orientation-only features versus orientation plus magnetometer; future studies should report both groupings explicitly.
  • The Lasso percentages are weights on feature groups, not causal effect sizes, so the actual error reduction from correcting any one factor may differ from its coefficient share because head distance, pose, and orientation are correlated with each other and with motion condition.
  • The head-distance proxy itself is a promising place for an independent check: adding a depth sensor or measured distance during the same tasks would test whether the top-ranked factor survives ground-truth measurement.
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 reports two user studies (N=10) in which participants performed everyday phone tasks under static and dynamic conditions (lying, sitting, standing, walking, and maze navigation), while the authors recorded synchronized IMU, camera, and gaze ground truth. Using a fine-tuned iTracker-based model with participant-specific SVR calibration, the authors measure 2D gaze-estimation RMSE across conditions and use Lasso regression to attribute error to sensor-derived factors. The stated findings are that dynamic conditions increase gaze error by up to 48.91% relative to static conditions, that head distance, head pose, and device orientation are the dominant error factors, and that one-time calibration (average RMSE 3.44 cm) is insufficient for mobile use, motivating adaptive calibration.

Significance. If the reported results hold, this is a useful empirical contribution to mobile gaze tracking: it provides a direct quantification of dynamic-vs-static accuracy degradation, demonstrates that task-switching behavior is more variable than within-task behavior, and offers a data-driven ranking of error factors. The study is comparatively unusual in collecting synchronized IMU + front-camera + gaze ground truth under semi-controlled motion conditions, and the core error measurements appear internally consistent with the reported ANOVA (F(4,36)=7.42, p<.001). However, the headline factor-attribution percentages are not currently trustworthy: the paper reports inconsistent numbers for the same quantities, and the top-ranked factor (head distance) is estimated from an unvalidated proxy. These issues must be resolved before the quantitative ranking can be accepted.

major comments (3)
  1. [Section 6.2.3; Section 1; Abstract] The factor-attribution shares are internally inconsistent across the paper. The abstract and the contribution list in Section 1 report head distance 35.87%, device orientation 32.29%, and head movements 10.18%. Section 6.2.3 reports, for the between-motion model, head distance 35.87%, head movements 10.19%, and device orientation 38.26%, and for the within-motion model, head distance 30.19%, head movements 16.25%, and device orientation 32.39%. The abstract's 32.29% value does not match either the between-motion (38.26%) or within-motion (32.39%) orientation share exactly, and it is paired with a between-motion head-distance share. Because these percentages are the paper's central quantitative contribution, the authors must specify which regression model supports each reported value and correct the inconsistent numbers.
  2. [Section 3.4, Eq. (1); Section 6.2.3] The top-ranked error factor, head distance, is estimated from a per-participant linear regression on the interpupillary-distance ratio (Eq. 1), calibrated with a ruler and never validated against ground-truth depth. The manuscript acknowledges in Section 3.4 that this estimate 'may introduce a minor source of measurement error,' but provides no quantification. Because the image-plane IPD depends on head pose, perspective foreshortening, and landmark detection noise—precisely the conditions varied in this study—the Lasso's assignment of roughly 36% (between-motion) or 30% (within-motion) of total coefficient weight to head distance could be an artifact of the proxy rather than a true causal contribution. The authors should validate Eq. (1) against a depth sensor (e.g., the iPhone's TrueDepth camera) across the experimental motion conditions, or at minimum report per-participant calibration residuals and a sensitivity analysis showing how the factor-share ranking changes under plausible proxy error.
  3. [Section 6.2.3] The Lasso 'contribution' percentages are computed from sums of coefficients, but the paper does not report the regression's fit quality (e.g., R²), the sign convention used (raw coefficients vs. absolute values), the cross-validated alpha selected by LassoCV, or the stability of selected features across cross-validation folds. Without these details, the reader cannot determine how much of the gaze error the 552-feature model actually explains, nor whether the factor ranking is robust. The authors should explicitly state whether the summed coefficients are absolute values, report the model R² and the number of nonzero coefficients, and provide a stability analysis (e.g., bootstrap or fold-wise feature selection).
minor comments (5)
  1. [Section 6.2.2] The sentence 'under dynamic conditions the error increases to 2.73 cm, dropping by nearly 48.91%' is self-contradictory; it should state that the error increases by nearly 48.91% from static to dynamic conditions.
  2. [Figure 14] The table embedded below Figure 14 contains formatting artifacts (e.g., '4.04333333') and mixes raw RMSE values with derived increase percentages; this material should be presented as a proper table with clear row and column labels.
  3. [Section 5.5.1] The statement that the head-distance filtering threshold of 100 cm follows from 'a lower one-sided confidence interval with 95% probability' is not self-explanatory; please describe the ergonomic rationale and the exact computation used.
  4. [Section 2.4] The phrase 'utilized Lasson regression process' contains a typo ('Lasson' should be 'Lasso'); please correct it.
  5. [Section 6.1.2] The claim that 'most participants exhibit ~4 clusters for lying and sitting and ~2 clusters for walking conditions' is only supported by box plots; reporting the clustering criterion (e.g., BIC or AIC) used for GMM model selection would strengthen the analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical error measurements and Lasso factor attribution are self-contained, and the IPD-based head-distance proxy is an input feature rather than a predicted target.

full rationale

The paper's core results are empirical: RMSE values under different motion conditions, comparisons of one-off, task-specific, and motion-specific calibration, and Lasso regression coefficients relating sensor features to measured gaze error. None of these assume the conclusion. The head-distance proxy in Eq. 1 is a per-participant linear fit from ruler-measured physical interpupillary distance and head-to-camera distance, but it serves only as an input feature to the Lasso; the target variable is the independently computed Euclidean distance between predicted and ground-truth gaze points. The headline 48.91% increase is computed from held-out 70% portions of test data after training on 30%, so it is not forced by construction. Self-citations to Lei et al. appear in related work and system design rationales but are not load-bearing for the central factor-identification result, and no uniqueness theorem is invoked. The acknowledged lack of ground-truth validation of the IPD-based distance proxy is a measurement-validity concern, not a circularity: an unvalidated covariate can bias a regression, but it does not make the derivation equivalent to its inputs. The paper is therefore self-contained against external benchmarks and the empirical claims are not circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims depend on a custom head-distance proxy fit per participant, unspecified filter and SVR parameters, and a lab-based approximation of real-world motion. These are not independently validated, and no external benchmarks are used for the factor attribution.

free parameters (6)
  • Head-distance regression coefficients beta0, beta1 (Eq. 1) = not reported
    Fitted per participant from manual ruler measurements of interpupillary distance and head-to-camera distance; the resulting head-distance estimate drives the top Lasso factor.
  • Head-distance filtering threshold = 100 cm
    Chosen as one-sided lower 95% confidence bound of the head-distance distribution; samples above are discarded as 'not looking at screen'.
  • SVR calibrator hyperparameters = not reported
    Kernel, C, epsilon, and feature scaling for the SVR calibrators are not specified; results may vary with these choices.
  • Butterworth low-pass filter parameters = not reported
    Order and cutoff frequency for IMU denoising are not stated.
  • Feature extraction window size = not reported
    The window over which IMU frequency, time, and cross-channel features are computed is not specified.
  • GMM cluster count selection = not reported
    The number of GMM components per participant is shown in box plots but the selection criterion (e.g., BIC or AIC) is not stated.
assumptions (5)
  • domain assumption Head-to-camera distance is a linear function of inverse interpupillary-distance ratio (Eq. 1)
    Pinhole proportionality between face size and distance; no depth-sensor validation, and the paper acknowledges possible measurement error (Section 3.4).
  • domain assumption Neighbor interpolation synchronizes 50 Hz IMU and 30 Hz camera timestamps accurately
    Section 5.5.1; interpolation error is not quantified and could bias sensor-error correlation.
  • domain assumption The iTracker-based model fine-tuned on RGBDGaze represents 2D gaze estimation methods generally
    Sections 3.1 and 7.3; no comparison to other 2D architectures or norm-based 3D methods, yet conclusions are framed broadly about 2D methods.
  • domain assumption Semi-controlled lab conditions (sofa, maze) approximate real-world mobile use
    Section 5.2 and 7.3; authors acknowledge limitation; motion patterns may differ in unconstrained settings.
  • domain assumption Lasso coefficient sums measure causal contribution shares
    Section 6.2.3; Lasso on collinear sensor features gives unstable attributions; no causal inference or confounding control.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying the Impact of Motion on 2D Gaze Estimation in Real-World Mobile Interactions." pith.science (2026). https://pith.science/paper/OXU2LPTB

@misc{pith2026250210570,
  author       = {Pith},
  title        = {Pith review of: Quantifying the Impact of Motion on 2D Gaze Estimation in Real-World Mobile Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OXU2LPTB}},
  note         = {Machine review of arXiv:2502.10570}
}
read the original abstract

Mobile gaze tracking involves inferring a user's gaze point or direction on a mobile device's screen from facial images captured by the device's front camera. While this technology inspires an increasing number of gaze-interaction applications, achieving consistent accuracy remains challenging due to dynamic user-device spatial relationships and varied motion conditions inherent in mobile contexts. This paper provides empirical evidence on how user mobility and behaviour affect mobile gaze tracking accuracy. We conduct two user studies collecting behaviour and gaze data under various motion conditions - from lying to maze navigation - and during different interaction tasks. Quantitative analysis has revealed behavioural regularities among daily tasks and identified head distance, head pose, and device orientation as key factors affecting accuracy, with errors increasing by up to 48.91% in dynamic conditions compared to static ones. These findings highlight the need for more robust, adaptive eye-tracking systems that account for head movements and device deflection to maintain accuracy across diverse mobile contexts.

Figures

Figures reproduced from arXiv: 2502.10570 by the authors.

Figure 1
Figure 1. Overview of the experimental mobile eye tracking system. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Head-to-camera distance calculation using interpupillary distance (IPD) regression. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Schematic overview of the user study design. User Study 1 explores spatial dynamics of mobile interactions [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Box plots of sensor data under each motion condition: [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: t-SNE plots of sensor features under each motion condition [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the number of motion clusters under different motion conditions [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: GMM clustering mapped to time-series sensor data on Participant 1 under static conditions [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: GMM clustering mapped to time-series sensor data on Participant 1 under dynamic conditions [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Box plots of motion data under each motion condition: [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: The left Figure 10a is an comparison overview of motion data in two user studies. The right Figure 10b [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Person-specific training and test regime of calibrations in different granularities for each motion condition [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Comparison of RMSE for within-task and between-task scenarios [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: t-SNE plots of motion data of calibration and test dataset under each motion condition [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Comparison of RMSE for within-motion and between-motion conditions [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Coordinate Singularities Break Conformal Coverage for Gaze and Head Pose

    cs.CV 2026-06 accept novelty 7.0 of 10

    Yaw–pitch and Euler conformal scores redistribute coverage near coordinate singularities; geodesic scores restore slice-conditional reliability without retraining.

  2. Factor-Informed Uncertainty Distillation for Gaze Estimation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    FIUD distills image-quality-based error predictions into a single-pass gaze uncertainty head, improving Spearman rank correlation and selective prediction versus ensembles, MC dropout, and heteroscedastic NLL baselines.

  3. MAC-Gaze: Motion-Aware Continual Calibration for Mobile Gaze Tracking

    cs.HC 2025-05 conditional novelty 6.0 of 10

    MAC-Gaze uses IMU-based motion detection and replay-based continual learning to automatically recalibrate mobile gaze trackers, reducing gaze estimation error by 19.9% on RGBDGaze and 31.7% on MotionGaze.

Reference graph

Works this paper leans on

61 extracted references · 60 canonical work pages · cited by 3 Pith papers

  1. [1]

    D., AND BAILEY , B

    ADAMCZYK , P. D., AND BAILEY , B. P. If not now, when? the effects of interruption at different moments within task execution. In Proceedings of the SIGCHI conference on Human factors in computing systems (2004), pp. 271–278

  2. [2]

    Enhancing accuracy and precision of eye tracker by head movement compensation and calibration

    ALEXIEV , K., T OSHKOV, T., AND DOJNOW , P. Enhancing accuracy and precision of eye tracker by head movement compensation and calibration. In Proceedings of the 20th International Conference on Computer Systems and Technologies (2019), pp. 226–233. 24 Quantifying Motion Impact on Mobile Gaze Tracking A PREPRINT

  3. [3]

    L., ET AL

    ANGUITA , D., G HIO , A., O NETO , L., P ARRA , X., R EYES -ORTIZ , J. L., ET AL . A public domain dataset for human activity recognition using smartphones. In Esann (2013), vol. 3, p. 3

  4. [4]

    Rgbdgaze: Gaze tracking on smartphones with rgb and depth data

    ARAKAWA , R., G OEL , M., H ARRISON , C., AND AHUJA , K. Rgbdgaze: Gaze tracking on smartphones with rgb and depth data. In Proceedings of the 2022 International Conference on Multimodal Interaction (2022), pp. 329–336

  5. [5]

    Accurate and robust eye contact detection during everyday mobile device interactions

    BÂCE , M., S TAAL, S., AND BULLING , A. Accurate and robust eye contact detection during everyday mobile device interactions. arXiv preprint arXiv:1907.11115 (2019)

  6. [6]

    Quantification of users’ visual attention during everyday mobile device interactions

    BÂCE , M., S TAAL, S., AND BULLING , A. Quantification of users’ visual attention during everyday mobile device interactions. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (2020), pp. 1–14

  7. [7]

    Efe: End-to-end frame-to-gaze estimation

    B ALIM , H., P ARK , S., W ANG , X., Z HANG , X., AND HILLIGES , O. Efe: End-to-end frame-to-gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2023), pp. 2687–2696

  8. [8]

    Adaptive feature fusion network for gaze tracking in mobile tablets

    BAO, Y., C HENG , Y., L IU, Y., AND LU, F. Adaptive feature fusion network for gaze tracking in mobile tablets. In 2020 25th International Conference on Pattern Recognition (ICPR)(2021), IEEE, pp. 9936–9943

Show all 61 references
  1. [9]

    Idiosyncratic feature-based gaze mapping

    B LIGNAUT , P. Idiosyncratic feature-based gaze mapping. Journal of Eye Movement Research 9, 3 (2016)

  2. [10]

    Gaze estimation using transformer

    CHENG , Y., AND LU, F. Gaze estimation using transformer. In 2022 26th International Conference on Pattern Recognition (ICPR) (2022), IEEE, pp. 3341–3347

  3. [11]

    Appearance-based gaze estimation with deep learning: A review and benchmark

    CHENG , Y., WANG , H., B AO, Y., AND LU, F. Appearance-based gaze estimation with deep learning: A review and benchmark. IEEE Transactions on Pattern Analysis and Machine Intelligence(2024)

  4. [12]

    Eye tracking in virtual reality

    CLAY, V., K ÖNIG , P., AND KOENIG , S. Eye tracking in virtual reality. Journal of eye movement research 12, 1 (2019)

  5. [13]

    Time- and space-efficient eye tracker calibration

    DREWES , H., P FEUFFER , K., AND ALT, F. Time- and space-efficient eye tracker calibration. In Proceedings of the 11th ACM Symposium on Eye Tracking Research & Applications(New York, NY , USA, 2019), ETRA ’19, Association for Computing Machinery

  6. [14]

    T., AND DUCHOWSKI , A

    DUCHOWSKI , A. T., AND DUCHOWSKI , A. T. Eye tracking methodology: Theory and practice. Springer, 2017

  7. [15]

    E., A TKESON , C

    FOGARTY , J., H UDSON , S. E., A TKESON , C. G., A VRAHAMI , D., F ORLIZZI , J., K IESLER , S., L EE, J. C., AND YANG , J. Predicting human interruptibility with sensors. ACM Transactions on Computer-Human Interaction (TOCHI) 12, 1 (2005), 119–146

  8. [16]

    A., M ONAY, F., AND ODOBEZ , J.-M

    FUNES MORA , K. A., M ONAY, F., AND ODOBEZ , J.-M. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the symposium on eye tracking research and applications (2014), pp. 255–258

  9. [17]

    Tutorials for our eye-tracking technology, April 6 2022

    G AZEPOINT . Tutorials for our eye-tracking technology, April 6 2022

  10. [18]

    Automatic gaze analysis: A survey of deep learning based approaches

    GHOSH , S., D HALL , A., H AYAT, M., K NIBBE , J., AND JI, Q. Automatic gaze analysis: A survey of deep learning based approaches. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 1 (2023), 61–84

  11. [19]

    Ml kit, March 2023

    G OOGLE LLC. Ml kit, March 2023

  12. [20]

    Flutter - build apps for any screen

    G OOGLE LLC. Flutter - build apps for any screen. https://flutter.dev, 2024. Accessed: 2023 Jan 07

  13. [21]

    W., AND JI, Q

    HANSEN , D. W., AND JI, Q. In the eye of the beholder: A survey of models for eyes and gaze. IEEE transactions on pattern analysis and machine intelligence 32, 3 (2009), 478–500

  14. [22]

    Gazeattentionnet: Gaze estimation with attentions

    HUANG , H., R EN, L., Y ANG , Z., Z HAN , Y., Z HANG , Q., AND LV, J. Gazeattentionnet: Gaze estimation with attentions. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2022), IEEE, pp. 2435–2439

  15. [23]

    X., AND BULLING , A

    HUANG , M. X., AND BULLING , A. Saccalib: Reducing calibration distortion for stationary eye trackers using saccadic eye movements. In Proceedings of the 11th ACM Symposium on Eye Tracking Research &amp; Applications (New York, NY , USA, 2019), ETRA ’19, Association for Comput...

  16. [24]

    X., L I, J., N GAI , G., AND LEONG , H

    HUANG , M. X., L I, J., N GAI , G., AND LEONG , H. V. Screenglint: Practical, in-situ gaze estimation on smartphones. In CHI ’17 (2017), ACM, p. 2546–2557

  17. [25]

    Tabletgaze: dataset and analysis for unconstrained appearance-based gaze estimation in mobile tablets

    HUANG , Q., V EERARAGHAVAN , A., AND SABHARWAL , A. Tabletgaze: dataset and analysis for unconstrained appearance-based gaze estimation in mobile tablets. Machine Vision and Applications 28, 5 (2017), 445–461

  18. [26]

    K., AND KO, J

    HUYNH , S., B ALAN , R. K., AND KO, J. imon: Appearance-based gaze tracking system on mobile devices. IMWUT 5, 4 (2021), 1–26. 25 Quantifying Motion Impact on Mobile Gaze Tracking A PREPRINT

  19. [27]

    Vads: Visual attention detection with a smartphone

    JIANG , Z., H AN, J., Q IAN , C., X I, W., Z HAO, K., D ING , H., T ANG , S., Z HAO, J., AND YANG , P. Vads: Visual attention detection with a smartphone. In IEEE INFOCOM 2016-The 35th Annual IEEE International Conference on Computer Communications (2016), IEEE, pp. 1–9

  20. [28]

    Rethinking model-based gaze estimation

    KAUR, H., J INDAL , S., AND MANDUCHI , R. Rethinking model-based gaze estimation. Proceedings of the ACM on computer graphics and interactive techniques 5, 2 (2022), 1–17

  21. [29]

    Gaze360: Physically unconstrained gaze estimation in the wild

    KELLNHOFER , P., R ECASENS , A., S TENT , S., M ATUSIK , W., AND TORRALBA , A. Gaze360: Physically unconstrained gaze estimation in the wild. In the IEEE/CVF International Conference on Computer Vision (2019), IEEE, pp. 6912–6921

  22. [30]

    The past, present, and future of gaze-enabled handheld mobile devices: Survey and lessons learned

    KHAMIS , M., A LT, F., AND BULLING , A. The past, present, and future of gaze-enabled handheld mobile devices: Survey and lessons learned. In Proceedings of the 20th International Conference on Human-Computer Interaction with Mobile Devices and Services (2018), pp. 1–17

  23. [31]

    Understanding face and eye visibility in front-facing cameras of smartphones used in the wild

    KHAMIS , M., B AIER , A., H ENZE , N., A LT, F., AND BULLING , A. Understanding face and eye visibility in front-facing cameras of smartphones used in the wild. In CHI ’18 (2018), ACM, p. 1–12

  24. [32]

    Eyemu interactions: Gaze+ imu gestures on mobile devices

    KONG , A., A HUJA , K., G OEL , M., AND HARRISON , C. Eyemu interactions: Gaze+ imu gestures on mobile devices. In Proceedings of the 2021 International Conference on Multimodal Interaction (2021), pp. 577–585

  25. [33]

    Eye tracking for everyone

    KRAFKA , K., K HOSLA , A., K ELLNHOFER , P., K ANNAN , H., B HANDARKAR , S., M ATUSIK , W., AND TORRALBA , A. Eye tracking for everyone. In Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 2176–2184

  26. [34]

    An end-to-end review of gaze estimation and its interactive applications on handheld mobile devices

    LEI, Y., H E, S., K HAMIS , M., AND YE, J. An end-to-end review of gaze estimation and its interactive applications on handheld mobile devices. ACM Comput. Surv. 56, 2 (sep 2023)

  27. [35]

    Dynamicread: Exploring robust gaze interaction methods for reading on handheld mobile devices under dynamic conditions

    LEI, Y., WANG , Y., C ASLIN , T., W ISOWATY, A., Z HU, X., K HAMIS , M., AND YE, J. Dynamicread: Exploring robust gaze interaction methods for reading on handheld mobile devices under dynamic conditions. Proceedings of the ACM on Human-Computer Interaction 7, ETRA23 (5 2023), 17

  28. [36]

    LEPOT, M., A UBIN , J.-B., AND CLEMENS , F. H. Interpolation in time series: An introductive overview of existing methods, their performance criteria and uncertainty assessment. Water 9, 10 (2017), 796

  29. [37]

    C., W ANG , Q., M EHTA, S., F OSTER , C., B ARNEY , E., S HAPIRO , L., V ENTOLA , P., AND SHIC , F

    LI, B., S NIDER , J. C., W ANG , Q., M EHTA, S., F OSTER , C., B ARNEY , E., S HAPIRO , L., V ENTOLA , P., AND SHIC , F. Calibration error prediction: Ensuring high-quality mobile eye-tracking. In 2022 Symposium on Eye Tracking Research and Applications(2022), pp. 1–7

  30. [38]

    3d model-based gaze tracking via iris features with a single camera and a single light source

    LIU, J., C HI, J., H U, W., AND WANG , Z. 3d model-based gaze tracking via iris features with a single camera and a single light source. IEEE Transactions on Human-Machine Systems 51, 2 (2020), 75–86

  31. [39]

    Gaze-dependent distortion correction for thick lenses in hmds

    MARTSCHINKE , J., M ARTSCHINKE , J., S TAMMINGER , M., AND BAUER , F. Gaze-dependent distortion correction for thick lenses in hmds. In 2019 IEEE Conference on Virtual Reality and 3D User Interfaces (VR) (2019), IEEE, pp. 1848–1851

  32. [40]

    Improving user experience of eye tracking-based interaction: Introspecting and adapting interfaces

    MENGES , R., K UMAR , C., AND STAAB, S. Improving user experience of eye tracking-based interaction: Introspecting and adapting interfaces. ACM Transactions on Computer-Human Interaction (TOCHI) 26, 6 (2019), 1–46

  33. [41]

    A., AND ESCALERA , S

    PALMERO , C., S ELVA, J., B AGHERI , M. A., AND ESCALERA , S. Recurrent cnn for 3d gaze estimation using appearance and shape cues. In 29th British Machine Vision Conference, BMVC 2018(2019)

  34. [42]

    V., AND TALATHI , S

    PALMERO , C., S HARMA , A., B EHRENDT , K., K RISHNAKUMAR , K., K OMOGORTSEV , O. V., AND TALATHI , S. S. Openeds2020 challenge on gaze tracking for vr: Dataset and results. Sensors 21, 14 (2021), 4769

  35. [43]

    D., M OLCHANOV , P., I QBAL , U., H ILLIGES , O., AND KAUTZ , J

    PARK , S., M ELLO , S. D., M OLCHANOV , P., I QBAL , U., H ILLIGES , O., AND KAUTZ , J. Few-shot adaptive gaze estimation. In Proceedings of the IEEE/CVF international conference on computer vision (2019), pp. 9368–9377

  36. [44]

    Pytorch: An imperative style, high-performance deep learning library

    PASZKE , A., G ROSS , S., M ASSA , F., L ERER , A., B RADBURY , J., C HANAN , G., K ILLEEN , T., L IN, Z., GIMELSHEIN , N., A NTIGA , L., ET AL . Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)

  37. [45]

    Scikit-learn: Machine learning in python

    PEDREGOSA , F., VAROQUAUX , G., G RAMFORT , A., M ICHEL , V., T HIRION , B., G RISEL , O., B LONDEL , M., PRETTENHOFER , P., W EISS , R., D UBOURG , V., ET AL . Scikit-learn: Machine learning in python. the Journal of machine Learning research 12 (2011), 2825–2830

  38. [46]

    PI, J., AND SHI, B. E. Task-embedded online eye-tracker calibration for improving robustness to head motion. In Proceedings of the 11th ACM Symposium on Eye Tracking Research & Applications(2019), pp. 1–9

  39. [47]

    Forecasting user attention during everyday mobile interactions using device-integrated and wearable sensors

    STEIL , J., M ÜLLER , P., S UGANO , Y., AND BULLING , A. Forecasting user attention during everyday mobile interactions using device-integrated and wearable sensors. In Proceedings of the 20th international conference on human-computer interaction with mobile devices and servi...

  40. [48]

    Learning-by-synthesis for appearance-based 3d gaze estimation

    SUGANO , Y., M ATSUSHITA , Y., AND SATO, Y. Learning-by-synthesis for appearance-based 3d gaze estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition(2014), pp. 1821–1828

  41. [49]

    An incremental learning method for uncon- strained gaze estimation

    SUGANO , Y., M ATSUSHITA , Y., S ATO, Y., AND KOIKE , H. An incremental learning method for uncon- strained gaze estimation. In Computer Vision – ECCV 2008 (Berlin, Heidelberg, 2008), D. Forsyth, P. Torr, and A. Zisserman, Eds., Springer Berlin Heidelberg, pp. 656–667

  42. [50]

    Revealing average screen time statistics, March 11 2024

    T EAM , B. Revealing average screen time statistics, March 11 2024

  43. [51]

    Get started with your eye tracker, April 24 2023

    T OBII . Get started with your eye tracker, April 24 2023

  44. [52]

    How to calibrate and validate in tobii pro lab, September 5 2022

    T OBII LTD. How to calibrate and validate in tobii pro lab, September 5 2022

  45. [53]

    Accelerating eye movement research via accurate and affordable smartphone eye tracking

    VALLIAPPAN , N., D AI, N., S TEINBERG , E., H E, J., R OGERS , K., R AMACHANDRAN , V., X U, P., S HO- JAEIZADEH , M., G UO, L., K OHLHOFF , K., ET AL . Accelerating eye movement research via accurate and affordable smartphone eye tracking. Nature communications 11, 1 (2020), 4553

  46. [54]

    Eyetab: Model-based gaze estimation on unmodified tablet computers

    WOOD , E., AND BULLING , A. Eyetab: Model-based gaze estimation on unmodified tablet computers. In Proceedings of the symposium on eye tracking research and applications(2014), pp. 207–210

  47. [55]

    Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation

    ZHANG , X., P ARK , S., B EELER , T., B RADLEY , D., T ANG , S., AND HILLIGES , O. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedi...

  48. [56]

    Revisiting data normalization for appearance-based gaze estimation

    ZHANG , X., S UGANO , Y., AND BULLING , A. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM Symposium on Eye Tracking Research & Applications(New York, NY , USA, 2018), ETRA ’18, Association for Computing Machinery

  49. [57]

    Revisiting data normalization for appearance-based gaze estimation

    ZHANG , X., S UGANO , Y., AND BULLING , A. Revisiting data normalization for appearance-based gaze estimation. In Proceedings of the 2018 ACM symposium on eye tracking research & applications(2018), pp. 1–9

  50. [58]

    Appearance-based gaze estimation in the wild

    ZHANG , X., S UGANO , Y., F RITZ , M., AND BULLING , A. Appearance-based gaze estimation in the wild. In CVPR ’15 (June 2015), IEEE, pp. 4511–4520

  51. [59]

    It’s written all over your face: Full-face appearance- based gaze estimation

    ZHANG , X., S UGANO , Y., F RITZ , M., AND BULLING , A. It’s written all over your face: Full-face appearance- based gaze estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (2017), pp. 51–60

  52. [60]

    Mpiigaze: Real-world dataset and deep appearance- based gaze estimation

    ZHANG , X., S UGANO , Y., F RITZ , M., AND BULLING , A. Mpiigaze: Real-world dataset and deep appearance- based gaze estimation. IEEE transactions on pattern analysis and machine intelligence 41, 1 (2017), 162–175

  53. [61]

    Mpiigaze: Real-world dataset and deep appearance- based gaze estimation

    ZHANG , X., S UGANO , Y., F RITZ , M., AND BULLING , A. Mpiigaze: Real-world dataset and deep appearance- based gaze estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 41, 1 (2019), 162–175. 27

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.