Pith. sign in

REVIEW 3 major objections 6 minor 40 references

EyeSeg: An Uncertainty-Aware Eye Segmentation Framework for AR/VR

T0 review · 3 major / 6 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 EyeSeg claims a provable uncertainty score for eye segmentation, and uses it to reject unreliable frames in AR/VR gaze estimation.

desk verdict Useful applied eye segmentation with a good uncertainty heuristic, but the 'provably' claim is not backed by the inference-time derivation. read the letter →

arxiv 2507.09649 v1 pith:FG5YGAKT submitted 2025-07-13 cs.CV

Zhengyuan Peng , Jianqing Xu , Shen Li , Jiazhen Ji , Yuge Huang , Jingyun Zhang , Jinmin Li , Shouhong Ding
show 3 more authors
Rizen Guo Xin Tan Lizhuang Ma
This is my paper
classification cs.CV
keywords eyesegmentationuncertaintyestimationaugmentedrealityvirtualgazeBayesiandeeplearningclosed-setpriorcovarianceprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the hard cases in AR/VR eye segmentation—motion blur, eyelid occlusion, and train-test domain gaps—can be treated as one problem: uncertainty. EyeSeg augments a standard segmentation network with a projection head that predicts a per-pixel covariance in latent space, trained by matching a Gaussian posterior to a closed-set prior centered at each class template. The paper's central theoretical result is that the trace of the optimal covariance equals the squared Euclidean distance between the latent feature and its class center, and it uses a statistic of the learned posterior, $s_{\mathrm{unc}} = \sum_{i,j} \ln \det(\Lambda_{x_{i,j}})$, as a whole-image uncertainty score. If this is right, AR/VR systems can reject or down-weight unreliable segmentations before gaze estimation, and the paper reports that filtering high-uncertainty images improves MIoU under occlusion, blur, and cross-domain settings.

What carries the argument

The load-bearing object is the uncertainty-aware projection head $\Lambda_\theta$, which predicts a diagonal covariance for every pixel. It is trained to approximate $(c_{y_{i,j}} - z_{i,j}) \odot (c_{y_{i,j}} - z_{i,j})$, the per-dimension squared distance from the latent feature to the class center derived from the classification weight matrix $W$. Theorem 1 identifies the trace of the optimal covariance with that squared distance, and the surrogate loss in Eq. (9) makes optimization tractable by avoiding the gradient-vanishing landscape of the original loss. At test time the same head outputs covariances that feed the uncertainty score $s_{\mathrm{unc}}$, which the framework uses to reject or reweight predictions.

What would settle it

On a held-out set with ground-truth labels, compute per-image $s_{\mathrm{unc}}$ and per-image segmentation error such as $1 - \mathrm{MIoU}$; if the rank correlation between them is not significantly positive, or if removing the highest-uncertainty images improves MIoU no more than removing random images, the claim that $s_{\mathrm{unc}}$ indicates segmentation uncertainty is refuted.

Watch

Extended reading notes

Core claim

EyeSeg's central claim is that uncertainty in eye segmentation can be learned explicitly rather than approximated by proxies. The theorem states that for the optimal diagonal covariance $\Lambda^*$ minimizing the cross-entropy loss between the Dirac prior $\delta(z_{i,j} - c_{y_{i,j}})$ and the Gaussian posterior $\mathcal{N}(z_{i,j}; f_\phi(x,i,j), \Lambda_\theta(x,i,j))$, one has $\operatorname{tr}(\Lambda^*_{x_{i,j}}) = \|z_{i,j} - c_{y_{i,j}}\|_2^2$. The paper then replaces the hard-to-optimize original loss with a surrogate that regresses the diagonal of $\Lambda$ onto the per-pixel squared error, and at inference derives $s_{\mathrm{unc}} = \sum_{i,j} \ln \det(\Lambda_{x_{i,j}})$ as a ranking score. Empirically, EyeSeg reports higher MIoU, E1, F1, and ACC than prior eye-segmentation methods and better retained-set MIoU than uncertainty baselines when the most uncertain images are removed.

Load-bearing premise

The test-time uncertainty score assumes the covariance predicted without knowing the true pixel label still behaves like the optimal covariance from training, but the paper's Theorem 1 is stated for the case where the true class center is available.

Editorial extensions

If this is right

  • A single scalar $s_{\mathrm{unc}}$ can rank eye frames by reliability, so gaze estimators can discard blurry, occluded, or out-of-domain frames before fusion.
  • Filtering the most uncertain fraction of test images raises the MIoU of the remaining set in the paper's experiments, which is the practical payoff of the uncertainty score.
  • Because the uncertainty head is a lightweight add-on (1.53G FLOPs total with the DenseElNet backbone), the approach stays within AR/VR real-time budgets.
  • The closed-set prior from eye-patch cropping is what makes the posterior tractable; the same recipe is claimed to generalize to the four-class eye structure of background, eye, iris, and pupil.

Reading between the lines

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

  • The trace identity is not eye-specific: any segmentation task with a fixed, known set of class template vectors could reuse the same surrogate loss, so the mechanism may transfer to other closed-set domains such as medical or industrial inspection.
  • The paper leaves unproven that the test-time covariance, computed without the true label, satisfies the same optimality condition as the training-time one; a direct check would be whether $\Lambda_\theta(x,i,j)$ tracks the true $(c_{y_{i,j}} - z_{i,j})^2$ on held-out labeled pixels.
  • One testable extension is to use $s_{\mathrm{unc}}$ as an active-learning acquisition function for collecting hard eye images, since the score appears to correlate with eye-semantics completeness.
  • The same score could serve as a calibration signal: ranking by $s_{\mathrm{unc}}$ should align with expected per-pixel error, a property the paper demonstrates indirectly through retained-set MIoU.
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 / 6 minor

Summary. EyeSeg proposes an uncertainty-aware eye segmentation framework for AR/VR. The pipeline first detects an eye patch, then runs a deterministic segmentation network f_phi whose logits are produced by a linear map W followed by softmax. A projection head Lambda_theta predicts a per-pixel diagonal covariance, and the model is trained so that this covariance approximates diag((c_y - z)^2), where c_y is the class center obtained from the pretrained rows of W and z is the deterministic feature. The paper claims that the trace of the optimal covariance equals the squared Euclidean distance to the class center (Theorem 1) and that, at inference, the sum of log-determinants of the predicted covariances is a provably grounded uncertainty score. Experiments on Else, Dikablis, LPW, and OpenEDS report improved segmentation MIoU and improved MIoU after removing high-uncertainty images, compared with several baselines.

Significance. If the theoretical claim were fully justified, the paper would offer a practical and interpretable uncertainty score for closed-set eye segmentation, with potential benefits for rejection and gaze-estimation fusion in AR/VR. The empirical evaluation covers multiple real-world datasets, reports both segmentation and uncertainty-rejection performance, and the code is publicly released. These are real strengths. However, the central 'provably' claim is currently not supported by the derivation: the test-time score used in Algorithm 2 is not shown to equal the negative log posterior density that the theory analyzes. The paper's contribution is therefore best assessed as a promising empirical heuristic with a suggestive but incomplete theoretical framing.

major comments (3)
  1. [Section 3.5, Eqs. (11)-(19)] The derivation of the uncertainty score silently replaces the learned posterior, defined in Section 3.3 as p_theta(z_ij|x) = N(z_ij; f_phi(x,i,j), Lambda_theta(x,i,j)), with a Gaussian centered at the true class center c_y and with covariance Lambda*. Equation (11) writes p(z_ij|x) with mean c_y and Lambda*, and Eq. (17) then uses the oracle identity gamma_ij = D. But Algorithm 2 computes s_unc from the network-predicted Lambda_theta(x,i,j) without access to c_y. No argument shows that the network-predicted covariance satisfies the same identity, especially for misclassified or out-of-distribution pixels. Consequently, the test-time score is not established to be -ln p(z|x), and the abstract's 'provably' claim is unsupported.
  2. [Theorem 1, Eq. (6), and Algorithm 2] Theorem 1 concerns the trace of the optimal per-pixel covariance, tr(Lambda*) = ||z - c_y||^2, while the deployed score in Eq. (19) and Algorithm 2 is sum_ij ln det(Lambda_theta(x,i,j)). For diagonal covariance matrices, trace and log-determinant are not monotonically related: for a fixed trace, log-determinant is maximized when the diagonal entries are as equal as possible and is minimized when one entry dominates. Thus even granting the training-time optimality result, the score actually used is not a monotone function of the theoretically justified statistic, so the link between the theorem and the deployed uncertainty score is missing.
  3. [Section 3.4, Eq. (10); Section 4.4, Table 2] The surrogate loss in Eq. (10) only enforces that Lambda_theta approximates diag((c_y - z)^2) on training pixels, where the true label is available. No generalization bound, calibration analysis, or empirical comparison of predicted covariances against the oracle Lambda* is provided to show that this approximation holds at test time, particularly for motion-blurred, occluded, or cross-domain images. The rejection experiments in Table 2 measure MIoU after filtering, but they do not test whether s_unc approximates the negative log posterior density. The theoretical grounding of the score therefore remains incomplete.
minor comments (6)
  1. [Section 3.3, Eq. (5)] The symbol z is used both for the random latent variable and for the deterministic feature f_phi(x,i,j); please distinguish these, since the Gaussian mean in p_theta is f_phi and the cross-entropy loss should be evaluated at z = c_y.
  2. [Abstract and Section 4] The abstract claims that EyeSeg 'empirically outperforms existing methods in downstream tasks, such as gaze estimation,' but Section 4 contains no gaze-estimation experiments; please add such experiments or qualify the claim.
  3. [Table 1] The FLOPs column mixes units without a consistent prefix (e.g., 125.85M, 1.53G, 3.5G); please standardize the units for readability.
  4. [Figures and cross-references] The text refers to Figure 4 both for segmentation results and for the uncertainty distribution; renumber the figures and update the cross-references to avoid confusion.
  5. [Section 4.1] The description of Else as 'we also curate a meticulously annotated dataset, Else [Fuhl et al., 2016b]' is ambiguous, since Else is an existing dataset; clarify whether new annotations are being contributed or whether the existing dataset is used.
  6. [References and text] There are encoding artifacts in the text and references, such as '´Ias' in the contributions list and '´Swirski' in the bibliography; please clean up the source files.

Circularity Check

1 steps flagged · score 5.0 of 10

The test-time uncertainty score is computed from the network's fitted covariance, while the theorem and posterior-density derivation apply only to a label-dependent oracle covariance; the proof is thereby separated from the deployed quantity.

  1. fitted input called prediction [Section 3.5, Eqs. (11)-(19), and Algorithm 2]
    "Substituting Eq. (17) into Eq. (11). we can derive the uncertainty estimate for the whole result from pixel-level estimate: ... Therefore, our proposed uncertainty score sunc for the whole segmented result is given by − ln p(z|x) ∝ X i,j ln det(Λ∗ xi,j ) = sunc (19) ... Λxi,j ← Λθ(x, i, j); sunc ← P i,j ln det(Λxi,j )."

    The derivation that makes sunc a principled negative-log-posterior statistic is performed for Λ*_xi,j, defined as the minimizer of the label-dependent loss L_ij (Eq. 5) with the true class center c_y, and the identity γ=D is obtained by setting Λ*=diag((c_y−z)^2) (Eqs. 12-17). At deployment, Algorithm 2 has no access to c_y and instead defines sunc from Λ_theta(x,i,j), the output of the variance head. The surrogate loss (Eq. 10) only trains Λ_theta to approximate diag((c_y−z)^2) on labeled training pixels; no theorem or generalization argument shows the predicted covariance equals the label-dependent oracle Λ* at test time.

full rationale

The paper's empirical evaluations, especially the uncertainty-based filtering experiments in Table 2, are self-contained against external benchmarks and provide some independent evidence that the proposed score has ranking value. However, the central theoretical framing is not fully self-consistent: Theorem 1 and the posterior-density derivation in Section 3.5 characterize an oracle covariance Λ* that depends on the true label c_y, while the deployed sunc is computed from the label-free network output Λ_theta. The surrogate loss trains Λ_theta to match the training-time feature-to-center squared distance, so sunc is essentially an aggregate of the model's own confidence target, and the claimed provable grounding does not transfer to the test-time quantity. There is also a mismatch between the theorem's trace statistic and the deployed log-determinant statistic, but the main circularity is the substitution of the fitted covariance for the oracle covariance. This is a partial, construction-level circularity rather than a fully forced self-citation chain; no load-bearing self-citations were found.

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

No new physical entities are introduced. The 'posterior' is a constructed statistical model, not a new entity. The central assumptions are the Gaussian latent-code model, the Dirac prior using classifier weights, and the unproven test-time extension of the optimality result.

assumptions (4)
  • domain assumption Latent codes z_i,j follow a Gaussian distribution with diagonal covariance parameterized by the projection head.
    Assumed in Section 3.3 (Eq. 4-5), not derived; it is a modeling choice that underpins the entire uncertainty formulation.
  • ad hoc to paper The prior over latent codes is a Dirac delta at the class center c_y = W^T e_y, where W is the pretrained segmentation classifier's weight matrix.
    This prior is chosen specifically to make the CE loss have a closed form and to tie uncertainty to the distance to the classifier's weight vectors; it is not a standard Bayesian prior and is introduced ad hoc.
  • domain assumption After eye-patch cropping, the label space is a closed set {0,1,2,3}, making the Dirac prior well-defined.
    This is stated in Section 3.1; it holds for the datasets considered but may not always hold in the wild, and the paper provides no evidence that the closed-set condition is critical.
  • ad hoc to paper The test-time uncertainty score s_unc = sum ln det(Lambda_theta(x,i,j)) is a valid proxy for predictive uncertainty even though the optimality condition in Theorem 1 requires the true label y_i,j.
    The paper does not prove this; it is the key leap from the training-time objective to the test-time use, and it is the weakest assumption in the framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EyeSeg: An Uncertainty-Aware Eye Segmentation Framework for AR/VR." pith.science (2026). https://pith.science/paper/FG5YGAKT

@misc{pith2026250709649,
  author       = {Pith},
  title        = {Pith review of: EyeSeg: An Uncertainty-Aware Eye Segmentation Framework for AR/VR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FG5YGAKT}},
  note         = {Machine review of arXiv:2507.09649}
}
read the original abstract

Human-machine interaction through augmented reality (AR) and virtual reality (VR) is increasingly prevalent, requiring accurate and efficient gaze estimation which hinges on the accuracy of eye segmentation to enable smooth user experiences. We introduce EyeSeg, a novel eye segmentation framework designed to overcome key challenges that existing approaches struggle with: motion blur, eyelid occlusion, and train-test domain gaps. In these situations, existing models struggle to extract robust features, leading to suboptimal performance. Noting that these challenges can be generally quantified by uncertainty, we design EyeSeg as an uncertainty-aware eye segmentation framework for AR/VR wherein we explicitly model the uncertainties by performing Bayesian uncertainty learning of a posterior under the closed set prior. Theoretically, we prove that a statistic of the learned posterior indicates segmentation uncertainty levels and empirically outperforms existing methods in downstream tasks, such as gaze estimation. EyeSeg outputs an uncertainty score and the segmentation result, weighting and fusing multiple gaze estimates for robustness, which proves to be effective especially under motion blur, eyelid occlusion and cross-domain challenges. Moreover, empirical results suggest that EyeSeg achieves segmentation improvements of MIoU, E1, F1, and ACC surpassing previous approaches. The code is publicly available at https://github.com/JethroPeng/EyeSeg.

Figures

Figures reproduced from arXiv: 2507.09649 by the authors.

Figure 1
Figure 1. The procedure of the proposed framework, EyeSeg. EyeSeg first detects eye regions, then extracts pixel-wise visual features and uncertainties used for Bayesian learning of a posterior which further yields an uncertainty score sunc during test. Here, τ denotes the threshold for decision-making. A detailed description of the process is provided in Algorithm 1 and Algorithm 2. tion over Union (MIoU), E1, F1, and ACC, w… view at source ↗
Figure 2
Figure 2. The function landscapes of the original loss Eq. (4) (left) and the surrogate loss Eq. (9) (right). For visualization, these loss functions are illustrated as functions of two free variables w1, w2 of the covariance Λxi,j . We mark an optimization path (P0 → P4) and the optimal point (OPT) for each. It can be observed that the original loss is prone to optimization issues such as gradient vanishing, whereas the surr… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Uncertainty distribution of different cross-domain datasets. The model is trained on the Else dataset and tested on oth￾ers datasets. Different domains have different distribution modes. Our uncertainty score correlates with eye semantics completeness. Compared to norm…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

  1. [1]

    Segnet: A deep convolu- tional encoder-decoder architecture for image segmenta- tion

    [Badrinarayanan et al., 2017] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolu- tional encoder-decoder architecture for image segmenta- tion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(12):2481–2495,

    work page 2017

  2. [5]

    Rethink- ing atrous convolution for semantic image segmentation

    [Chen et al., 2017b] Liang-Chieh Chen, George Papan- dreou, Florian Schroff, and Hartwig Adam. Rethink- ing atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587,

    arXiv

  3. [6]

    Encoder-decoder with atrous separable convolution for se- mantic image segmentation

    [Chen et al., 2018] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for se- mantic image segmentation. In Proceedings of the Eu- ropean Conference on Computer Vision , pages 801–818,

    work page 2018

  4. [10]

    Real-time gaze tracking with event-driven eye segmentation

    [Feng et al., 2022] Yu Feng, Nathan Goulding-Hotta, Asif Khan, Hans Reyserhove, and Yuhao Zhu. Real-time gaze tracking with event-driven eye segmentation. In 2022 IEEE Conference on Virtual Reality and 3D User Inter- faces, pages 399–408. IEEE,

    work page 2022

  5. [14]

    Teyed: Over 20 million real-world eye images with pupil, eyelid, and iris 2d and 3d segmen- tations, 2d and 3d landmarks, 3d eyeball, gaze vector, and eye movement types

    [Fuhl et al., 2021] Wolfgang Fuhl, Gjergji Kasneci, and Enkelejda Kasneci. Teyed: Over 20 million real-world eye images with pupil, eyelid, and iris 2d and 3d segmen- tations, 2d and 3d landmarks, 3d eyeball, gaze vector, and eye movement types. In IEEE International Symposium on Mixed and Augmented Reality , pages 367–375. IEEE,

    work page 2021

  6. [15]

    Teyed: Over 20 million real-world eye images with pupil, eyelid, and iris 2d and 3d segmen- tations, 2d and 3d landmarks, 3d eyeball, gaze vector, and eye movement types,

    [Fuhl et al., 2022] Wolfgang Fuhl, Gjergji Kasneci, and Enkelejda Kasneci. Teyed: Over 20 million real-world eye images with pupil, eyelid, and iris 2d and 3d segmen- tations, 2d and 3d landmarks, 3d eyeball, gaze vector, and eye movement types,

    work page 2022

  7. [16]

    Pistol: Pupil invisible supportive tool in the wild

    [Fuhl et al., 2024] Wolfgang Fuhl, Daniel Weber, and Shahram Eivazi. Pistol: Pupil invisible supportive tool in the wild. SN Computer Science, 5(3):276,

    work page 2024

  8. [17]

    Garbin, Yiru Shen, Immo Schuetz, Robert Cavin, Gregory Hughes, and Sachin S

    [Garbin et al., 2019] Stephan J. Garbin, Yiru Shen, Immo Schuetz, Robert Cavin, Gregory Hughes, and Sachin S. Ta- lathi. Openeds: Open eye dataset. CoRR, abs/1905.03702,

    arXiv 2019

Show all 40 references
  1. [18]

    Dataset for eye tracking on a virtual reality platform

    [Garbin et al., 2020] Stephan Joachim Garbin, Oleg Ko- mogortsev, Robert Cavin, Gregory Hughes, Yiru Shen, Immo Schuetz, and Sachin S Talathi. Dataset for eye tracking on a virtual reality platform. In ACM Symposium on Eye Tracking Research and Applications , pages 1–10,

    2020

  2. [20]

    Eye tracking in the wild

    [Hansen and Pece, 2005] Dan Witzner Hansen and Arthur EC Pece. Eye tracking in the wild. Com- puter Vision and Image Understanding , 98(1):155–181,

    2005

  3. [22]

    What uncertainties do we need in bayesian deep learning for computer vision? Neural Information Processing Systems, 30,

    [Kendall and Gal, 2017] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Neural Information Processing Systems, 30,

    2017

  4. [23]

    Ellseg: An ellipse segmentation framework for robust gaze tracking

    [Kothari et al., 2021] Rakshit S Kothari, Aayush K Chaud- hary, Reynold J Bailey, Jeff B Pelz, and Gabriel J Diaz. Ellseg: An ellipse segmentation framework for robust gaze tracking. IEEE Transactions on Visualization and Com- puter Graphics, 27(5):2757–2767,

    2021

  5. [24]

    Dudes: Deep uncertainty distillation using ensembles for semantic seg- mentation

    [Landgraf et al., 2024] Steven Landgraf, Kira Wursthorn, Markus Hillemann, and Markus Ulrich. Dudes: Deep uncertainty distillation using ensembles for semantic seg- mentation. PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science , 92(2):101–114,

    2024

  6. [25]

    Primitivenet: decomposing the global constraints for referring segmentation

    [Liu et al., 2024] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2(1):16,

    2024

  7. [26]

    Fully convolutional networks for seman- tic segmentation

    [Long et al., 2015] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for seman- tic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431– 3440,

    2015

  8. [27]

    Shape constrained net- work for eye segmentation in the wild

    [Luo et al., 2020] Bingnan Luo, Jie Shen, Shiyang Cheng, Yujiang Wang, and Maja Pantic. Shape constrained net- work for eye segmentation in the wild. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 1952–1960,

    2020

  9. [28]

    Generalized category discovery in semantic segmentation

    [Peng et al., 2023] Zhengyuan Peng, Qijian Tian, Jianqing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, and Lizhuang Ma. Generalized category discovery in semantic segmentation. arXiv preprint arXiv:2311.11525,

    2023 arXiv

  10. [29]

    Yolov3: An incremental improvement

    [Redmon and Farhadi, 2018] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv,

    2018

  11. [30]

    U-net: Convolutional networks for biomedical image segmentation

    [Ronneberger et al., 2015] Olaf Ronneberger, Philipp Fis- cher, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. Medical Image Com- puting and Computer Assisted Intervention,

    2015

  12. [31]

    Uncertainty-aware panop- tic segmentation

    [Sirohi et al., 2023] Kshitij Sirohi, Sajad Marvi, Daniel B¨uscher, and Wolfram Burgard. Uncertainty-aware panop- tic segmentation. IEEE Robotics and Automation Letters, 8(5):2629–2636,

    2023

  13. [32]

    Robust real-time pupil tracking in highly off-axis images

    [´Swirski et al., 2012] Lech ´Swirski, Andreas Bulling, and Neil Dodgson. Robust real-time pupil tracking in highly off-axis images. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 173–176,

    2012

  14. [34]

    Labeled pupils in the wild: A dataset for studying pupil detection in unconstrained en- vironments

    [Tonsen et al., 2015] Marc Tonsen, Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Labeled pupils in the wild: A dataset for studying pupil detection in unconstrained en- vironments. CoRR, abs/1511.05768,

    2015 arXiv

  15. [35]

    Labelled pupils in the wild: a dataset for studying pupil detection in unconstrained en- vironments

    [Tonsen et al., 2016] Marc Tonsen, Xucong Zhang, Yusuke Sugano, and Andreas Bulling. Labelled pupils in the wild: a dataset for studying pupil detection in unconstrained en- vironments. In Proceedings of the Ninth Biennial ACM Symposium on Eye Tracking Research & Applications ...

    2016

  16. [36]

    Towards complete and accurate iris segmentation using deep multi- task attention network for non-cooperative iris recognition

    [Wang et al., 2020] Caiyong Wang, Jawad Muhammad, Yunlong Wang, Zhaofeng He, and Zhenan Sun. Towards complete and accurate iris segmentation using deep multi- task attention network for non-cooperative iris recognition. IEEE Transactions on information forensics and security, ...

    2020

  17. [37]

    To- ward accurate and reliable iris segmentation using uncer- tainty learning

    [Wei et al., 2021] Jianze Wei, Huaibo Huang, Muyi Sun, Yunlong Wang, Min Ren, Ran He, and Zhenan Sun. To- ward accurate and reliable iris segmentation using uncer- tainty learning. arXiv preprint arXiv:2110.10334,

    2021 arXiv

  18. [38]

    Seg- former: Simple and efficient design for semantic segmen- tation with transformers

    [Xie et al., 2021] Enze Xie, Wenhai Wang, Zhiding Yu, An- ima Anandkumar, Jose M Alvarez, and Ping Luo. Seg- former: Simple and efficient design for semantic segmen- tation with transformers. Advances in neural information processing systems, 34:12077–12090,

    2021

  19. [39]

    Deepvog: Open-source pupil segmentation and gaze estimation in neuroscience using deep learning

    [Yiu et al., 2019] Yuk-Hoi Yiu, Moustafa Aboulatta, Theresa Raiser, Leoni Ophey, Virginia L Flanagin, Peter zu Eulenburg, and Seyed-Ahmad Ahmadi. Deepvog: Open-source pupil segmentation and gaze estimation in neuroscience using deep learning. Journal of Neuro- science Methods,

    2019

  20. [40]

    A full-set tooth segmentation model based on improved pointnet++

    [Yuan et al., 2023] Li Yuan, Xinyi Liu, Jiannan Yu, and Yan- feng Li. A full-set tooth segmentation model based on improved pointnet++. Visual Intelligence, 1(1):21, 2023

    2023

  21. [2005]

    Sipformer: Seg- mentation of multiocular biometric traits with transform- ers

    [Hassan et al., 2022] Bilal Hassan, Taimur Hassan, Ramsha Ahmed, Naoufel Werghi, and Jorge Dias. Sipformer: Seg- mentation of multiocular biometric traits with transform- ers. IEEE Transactions on Instrumentation and Measure- ment, 72:1–14,

    2022

  22. [2012]

    Positive-negative recep- tive field reasoning for omni-supervised 3d segmentation

    [Tan et al., 2023] Xin Tan, Qihang Ma, Jingyu Gong, Ji- achen Xu, Zhizhong Zhang, Haichuan Song, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Positive-negative recep- tive field reasoning for omni-supervised 3d segmentation. IEEE Transactions on Pattern Analysis and Machine In- telli...

    2023

  23. [2015]

    Pupilnet: Convolutional neural networks for robust pupil detection

    [Fuhl et al., 2016a] Wolfgang Fuhl, Thiago Santini, Gjergji Kasneci, and Enkelejda Kasneci. Pupilnet: Convolutional neural networks for robust pupil detection. arXiv preprint arXiv:1601.04902,

    arXiv

  24. [2016]

    Pupilnet v2

    [Fuhl et al., 2017] Wolfgang Fuhl, Thiago Santini, Gjergji Kasneci, Wolfgang Rosenstiel, and Enkelejda Kasneci. Pupilnet v2. 0: Convolutional neural networks for cpu based real time robust pupil detection. arXiv preprint arXiv:1711.00112,

    2017 arXiv

  25. [2017]

    A framework for generalizable neural networks for robust estimation of eyelids and pupils

    [Biswas and Lescroart, 2023] Arnab Biswas and Mark D Le- scroart. A framework for generalizable neural networks for robust estimation of eyelids and pupils. Behavior Re- search Methods, pages 1–23,

    2023

  26. [2018]

    Beyond the label itself: La- tent labels enhance semi-supervised point cloud panoptic segmentation

    [Chen et al., 2024] Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, and Yuan Xie. Beyond the label itself: La- tent labels enhance semi-supervised point cloud panoptic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1245–1253,

    2024

  27. [2019]

    Se- mantic image segmentation with deep convolutional nets and fully connected crfs

    [Chen et al., 2014] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Se- mantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062,

    2014 arXiv

  28. [2020]

    Boundary-aware geometric encoding for semantic seg- mentation of point clouds

    [Gong et al., 2021] Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric encoding for semantic seg- mentation of point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 1424–1432,

    2021

  29. [2021]

    Sernet-former: Semantic seg- mentation by efficient residual network with attention- boosting gates and attention-fusion networks

    [Erisen, 2024] Serdar Erisen. Sernet-former: Semantic seg- mentation by efficient residual network with attention- boosting gates and attention-fusion networks. arXiv preprint arXiv:2401.15741,

    2024

  30. [2022]

    Ex- cuse: Robust pupil detection in real-world scenarios

    [Fuhl et al., 2015] Wolfgang Fuhl, Thomas K ¨ubler, Katrin Sippel, Wolfgang Rosenstiel, and Enkelejda Kasneci. Ex- cuse: Robust pupil detection in real-world scenarios. In Computer Analysis of Images and Patterns , pages 39–51. Springer,

    2015

  31. [2023]

    Ritnet: Real-time semantic segmentation of the eye for gaze tracking

    [Chaudhary et al., 2019] Aayush K Chaudhary, Rakshit Kothari, Manoj Acharya, Shusil Dangi, Nitinraj Nair, Reynold Bailey, Christopher Kanan, Gabriel Diaz, and Jeff B Pelz. Ritnet: Real-time semantic segmentation of the eye for gaze tracking. In 2019 IEEE/CVF Interna- tional Co...

    2019

  32. [2024]

    [Czolbe et al., 2021] Steffen Czolbe, Kasra Arnavaz, Oswin Krause, and Aasa Feragen. Is segmentation uncertainty useful? In Information Processing in Medical Imaging: 27th International Conference, IPMI 2021, Virtual Event, June 28–June 30, 2021, Proceedings 27 , pages 715–726...

    2021

Pith tools

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