Pith. sign in

REVIEW 3 major objections 7 minor 43 references

EM-Net: Gaze Estimation with Expectation Maximization Algorithm

T0 review · 3 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read This paper proposes EM-Net, a lightweight gaze estimator that combines a global attention mechanism with an expectation-maximization module, and claims it beats the GazeNAS-ETH baseline on three public benchmarks using only half the…

desk verdict A plausible architecture with an uncontrolled head-to-head: the 50%-data claim rests on published baselines not retrained under the same protocol. read the letter →

arxiv 2412.08074 v1 pith:B2IWUKI2 submitted 2024-12-11 cs.CV cs.LG

classification cs.CVcs.LG
keywords gazeestimationexpectationmaximizationglobalattentionmechanismlightweightnetworkshifted-windowself-attentiondataefficiencyMPIIFaceGaze360
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 show that a lightweight gaze-estimation model can be accurate without large datasets or large models. EM-Net is built on a slimmed MobileNetV3 backbone; a proposed Global Attention Mechanism (GAM) uses shifted-window self-attention to enlarge the receptive field and fuse spatial and channel information, while an Expectation-Maximization (EM) module refines the CNN features by iteratively estimating latent structure. The paper's central experimental claim is that, trained on only 50% of the training samples of MPIIFaceGaze, Gaze360, and RT-Gene, EM-Net reduces angular error relative to GazeNAS-ETH by 2.02%, 2.2%, and 2.03%, respectively. If this holds, accurate gaze estimation becomes cheaper to train and easier to deploy on resource-limited devices.

What carries the argument

The load-bearing machinery is the pairing of a Global Attention Mechanism (GAM) with an Expectation-Maximization module inserted between a MobileNetV3-derived backbone and gaze regression. GAM reuses the shifted-window and masking idea from Swin Transformer to exchange information across spatial windows and across channels, explicitly enlarging the effective receptive field. The EM module first projects the 960-channel feature map down to 540 channels with a 1x1 convolution, then iterates the classic E-step/M-step update on the feature distribution: it computes the expected log-likelihood over hidden variables given current parameters, then re-estimates parameters by maximization. This iteration is what the paper claims lets the network deal with occluded or noisy facial regions and generalize from fewer samples.

What would settle it

Re-run GazeNAS-ETH under exactly EM-Net's training protocol—the same odd-indexed 50% split, the same ETH-XGaze initialization, the same preprocessing—and evaluate with the same angular-error metric. If GazeNAS-ETH's error then equals or beats EM-Net's on MPIIFaceGaze, Gaze360, or RT-Gene, the paper's central comparison collapses.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a carefully combined attention-and-EM architecture can maintain or improve gaze accuracy while halving the training data. EM-Net reports angular errors of 3.88° on MPIIFaceGaze, 10.29° on Gaze360, and 6.27° on RT-Gene, improvements of 2.02%, 2.2%, and 2.03% over GazeNAS-ETH, with 2.93M parameters and 0.31 GFLOPs. The authors attribute the gains to GAM, which captures long-distance dependencies through shifted-window multi-head self-attention and masks, and to the EM module, which iterates between an expectation step and a maximization step over the feature map so that occluded or noisy input regions are handled statistically. In ablation, removing either GAM or the EM module costs a similar amount of accuracy, and removing both raises the Gaze360 error by 0.74° (7.2%).

Load-bearing premise

The central claim depends on the assumption that the GazeNAS-ETH numbers quoted from its original paper are directly comparable to EM-Net's results, even though EM-Net trains on a different 50% subsample, uses ETH-XGaze pretraining, and applies its own preprocessing.

Editorial extensions

If this is right

  • Deploying gaze estimation on edge devices becomes more feasible: EM-Net's 2.93M parameters and 0.31 GFLOPs are an order of magnitude smaller than many transformer-based competitors while maintaining competitive accuracy.
  • Training cost drops: using half the training data of three standard benchmarks directly halves annotation and processing effort for this task.
  • The EM module's robustness to Gaussian noise (angle errors degrade only roughly 0.2–0.7° on MPIIFaceGaze and 0.4–0.9° on Gaze360) suggests the architecture is suited to real-world images with sensor noise or occlusion.
  • The comparison implies that receptive-field expansion (via GAM) and latent-feature refinement (via EM) are roughly equally important, since ablating either one costs a similar amount of accuracy.

Reading between the lines

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

  • Because the reported margins over GazeNAS-ETH are only 0.08° to 0.24°, a controlled re-run under identical training splits, pretraining, and preprocessing could easily reverse the ranking; the 'data efficiency' conclusion would be stronger if the baselines had been retrained on the same 50% subset.
  • The EM module, as implemented, is a feature-refinement block rather than a full generative model fit; ablating the number of EM iterations (including a zero-iteration pass) would show whether the improvement comes from the EM principle or simply from the extra computation.
  • If the generalization benefit is real, it should become more pronounced as the training fraction drops; testing at 25% and 10% data would provide a direct, falsifiable extension.
  • GAM borrows shifted-window self-attention from the Swin design, so the comparison against SE and CBAM mainly shows that some form of windowed self-attention helps; swapping GAM for a standard Swin attention module at equal FLOPs would isolate what the GAM-specific arrangement adds.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes EM-Net, a lightweight gaze estimation network built on an improved MobileNetV3 backbone and augmented with a Global Attention Mechanism (GAM) and an Expectation Maximization (EM) module. The authors claim that EM-Net achieves lower angular errors than GazeNAS-ETH and other baselines on MPIIFaceGaze, Gaze360, and RT-Gene while using only 50% of the training data, and that it is robust to Gaussian noise. The experiments report single-run angular errors, parameter counts, FLOPs, and inference times, along with ablation studies and attention-replacement comparisons.

Significance. If the central claims were properly validated, the paper would contribute a lightweight gaze estimation model with a notably low parameter count (2.93M) and computational cost (0.31 GFLOPs), and would demonstrate an interesting data-efficiency property via the EM module. However, the current evidence is not convincing: the headline comparison against published baseline numbers is uncontrolled, the EM module is described only at the level of generic EM equations without a concrete instantiation, and all quantitative results lack error bars or repeated-seed runs. The ablation and attention comparisons show plausible trends, but the small differences (as low as 0.08°) could easily be within run-to-run variance.

major comments (3)
  1. [Section 4.3, Table 2] The central claim of a 0.08°–0.24° improvement over GazeNAS-ETH using half the training data is based on baseline angular errors cited from the original papers, not from re-running those models under EM-Net's protocol. EM-Net differs in two material ways: it uses only odd-indexed 50% of the training samples (§4.2) and it initializes from ETH-XGaze pretraining (§4.1). Without a same-protocol re-run of GazeNAS-ETH (or at least training GazeNAS-ETH on the same 50% subset with the same pretraining), the comparison is not controlled, and the claimed margins are small enough that they could be explained by training noise or the pretraining initialization. The '50% data' advantage is therefore unverified.
  2. [Section 3.2, Algorithm 1] The EM module is not actually specified as a network layer. Algorithm 1 states the textbook EM update equations for a generic probabilistic model P(Y,Z|θ), but the paper does not define what Y, Z, and θ are in the context of the feature maps f∈R^{4×4×540}, nor how the E-step and M-step are computed on those tensors. The line 'θ(i) = Conv1×1' is not a meaningful initialization of model parameters for a probability model. Consequently, the module is not reproducible, and the assertion that it 'retains the characteristic that the EM algorithm is good at processing unobserved variables' (§4.3) has no concrete algorithmic content.
  3. [Sections 4.3, 4.4, 4.6 (Tables 2, 4, 6)] All angular errors in the paper are reported without standard deviations, confidence intervals, or the number of repeated runs. This is particularly problematic because the central comparison yields differences as small as 0.08° (MPIIFaceGaze) and the GAM-vs-SE/CBAM and ablation differences in Tables 4 and 6 are 0.10°–0.74°. Without error bars, one cannot distinguish the proposed method's reported improvements from run-to-run variation, especially given the known variance of gaze estimation results across training runs.
minor comments (7)
  1. [Algorithm 1] In the E-step, Q(θ, θ(i+1)) should be Q(θ, θ(i)); the same index correction applies to the maximization step.
  2. [Section 4.3] The percentage improvement for Gaze360 is reported as 2.2%, but 0.24° relative to 10.52° is approximately 2.28%; the rounding should be made consistent.
  3. [Section 4.2] Clarify whether the odd-indexed sampling is applied only to the training split or to the entire dataset, and specify how the test set is constructed; this affects the comparability of the reported errors.
  4. [Section 4.5] State explicitly whether Gaussian noise is added to the test images as well as the training images in the robustness experiments, and whether the same noise level is used for both.
  5. [Section 3.1.2] The GAM appears to combine SE-like channel attention with Swin Transformer's shifted-window self-attention; the novelty relative to the Swin Transformer block should be discussed, and a direct comparison with a standard Swin block would be more informative than the SE/CBAM comparison alone.
  6. [Table 3] The text states that inference times are computed under the same hardware conditions, but it is unclear whether the baseline times are re-measured or taken from their original papers; the subsequent sentence saying parameters and FLOPs are cited from original papers adds ambiguity.
  7. [General] The paper contains numerous English grammar and missing-article errors (e.g., 'the proposed Global Attention Mechanism(GAM) is added to extract features related to gaze estimation to improve the model's ability'); a careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EM-Net's accuracy and data-efficiency claims are empirical comparisons against external published baselines, not reductions of the paper's own equations or self-citations.

full rationale

The paper's claimed contributions are architectural (GAM, EM module, lightweight EM-Net) and empirical (angular error, parameters, FLOPs, inference time, robustness). None of the central claims is defined in terms of the target result. The EM module in Algorithm 1 restates the standard expectation-maximization update for a latent-variable model and is used as a feature-refinement layer; the final gaze output is produced by a learned MLP regression (Eq. 3), not by solving the EM equations for the evaluation metric. The comparison in Table 2 uses angular errors from original papers, while EM-Net is trained on 50% odd-indexed samples with ETH-XGaze pretraining; this is a protocol-comparability weakness that could affect the validity of the reported margins, but it is not circularity because the EM-Net numbers are independently measured on held-out data and are not constructed from the baseline numbers. No load-bearing self-citation was found: references to prior EM routing work (e.g., [27]) are background, not justification of this paper's results. The ablation and attention-comparison experiments are self-consistent empirical evaluations. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The ledger lists the unspecified or hand-chosen ingredients the central claim depends on: the unstated EM module implementation, the arbitrary channel dimension and iteration count, the odd-index 50% subset, and the assumption that published baseline errors are protocol-comparable.

free parameters (3)
  • EM module iteration count = Not reported
    Algorithm 1 loops with no stopping criterion or iteration count; latency and the claimed feature refinement both depend on it.
  • Conv1x1 output channels before EM module = 540
    Section 3.2 reduces from 960 to 540 channels without a sensitivity study; this dimension is part of the reported 0.31 GFLOPs.
  • 50% training subset parity = odd serial numbers
    Section 4.2 selects odd-indexed samples from each dataset; no justification or alternative subset check is given.
assumptions (4)
  • ad hoc to paper The textbook EM update in Algorithm 1 can be instantiated as the feature-refinement module used in EM-Net.
    Section 3.2 and Algorithm 1 give only generic Q(theta, theta') and argmax updates; the observed variable, hidden variable, parameters, and probability model are never mapped to network features or activations.
  • domain assumption Baseline angular errors in Table 2 are comparable to EM-Net's errors under the same evaluation setup.
    Section 4.3 copies baseline numbers from original papers; no baseline is retrained on 50% interval-sampled data or with ETH-XGaze initialization.
  • domain assumption Odd-serial-number interval sampling yields a representative 50% training subset.
    Section 4.2 introduces the sampling rule without analyzing subject, pose, illumination, or temporal distribution of the retained frames.
  • domain assumption Pretraining on ETH-XGaze transfers to MPIIFaceGaze, Gaze360, and RT-Gene.
    Section 4.1 initializes EM-Net with ETH-XGaze pretrained weights; the cited baselines may or may not use the same pretraining, so the comparison includes an unstated transfer advantage.
invented entities (2)
  • EM module
    purpose: Feature refinement after the backbone; claimed to handle occlusion and noise and to improve generalization (Section 3.2).
    No concrete EM update equations, no code, and no independent task validation; the only evidence is an ablation within this paper.
  • Global Attention Mechanism (GAM)
    purpose: Combines channel gating with shifted-window multi-head self-attention to expand the receptive field (Section 3.1.2).
    It is a composition of SE and Swin-style attention; the paper's own ablation and visualization are the only support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EM-Net: Gaze Estimation with Expectation Maximization Algorithm." pith.science (2026). https://pith.science/paper/B2IWUKI2

@misc{pith2026241208074,
  author       = {Pith},
  title        = {Pith review of: EM-Net: Gaze Estimation with Expectation Maximization Algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2IWUKI2}},
  note         = {Machine review of arXiv:2412.08074}
}
read the original abstract

In recent years, the accuracy of gaze estimation techniques has gradually improved, but existing methods often rely on large datasets or large models to improve performance, which leads to high demands on computational resources. In terms of this issue, this paper proposes a lightweight gaze estimation model EM-Net based on deep learning and traditional machine learning algorithms Expectation Maximization algorithm. First, the proposed Global Attention Mechanism(GAM) is added to extract features related to gaze estimation to improve the model's ability to capture global dependencies and thus improve its performance. Second, by learning hierarchical feature representations through the EM module, the model has strong generalization ability, which reduces the need for sample size. Experiments have confirmed that, on the premise of using only 50% of the training data, EM-Net improves the performance of Gaze360, MPIIFaceGaze, and RT-Gene datasets by 2.2%, 2.02%, and 2.03%, respectively, compared with GazeNAS-ETH. It also shows good robustness in the face of Gaussian noise interference.

Figures

Figures reproduced from arXiv: 2412.08074 by the authors.

Figure 1
Figure 1. EM-Net. to achieve significant results in multiple publicly available datasets, enhancing resistance to white-box adversarial at￾tacks. Zhang et al. [27] propose a lightweight Multi Inter￾est Retrieval Network (MIRN) by introducing EM routing, which adaptively learns vector-guided model training of user representations and achieves significant performance in recommendation systems. In the task of gaze estimation, th… view at source ↗
Figure 2
Figure 2. Bneck network structure. NL indicates the activation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. SE and GAM. h − swish[x] = x ReLU6(x + 3) 6 (2) 3.1.2 Global attention mechanism(GAM) SE [24] attention can make the model focus more on useful channel information by learning adaptive channel weights. However, SE attention only fuses spatial infor￾mation through global average pooling, which makes it unable to capture attention in the spatial dimension and does not fully utilize contextual information. In response … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Schematic diagram of GAM information exchange. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: EM Module. horizontal yaw angle and the vertical pitch angle. The cal￾culation formula is as follows: (yaw, pitch) = MLP(GAP(f)) (3) Where f represents the features input into the Gaze Regres￾sion, GAP represents the global average pooling, and MLP represents the multi…
Figure 6
Figure 6. Figure 6: Visualization of receptive field. the data, provide maximum likelihood estimation of parameters, and increase the likelihood function value of the data every iteration, ensuring the convergence of the model. When dealing with incomplete data, the model parameters are u…
Figure 7
Figure 7. Figure 7: Example of data with Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: Visualization of gaze estimation direction. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 8
Figure 8. Figure 8: Example of the gaze360 dataset. 4.7. Data visualization In order to provide a more intuitive representation of the gaze estimation results of the EM-Net model on three datasets, [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Appearance-based gaze estimation for asd diagnosis,

    J. Li, Z. Chen, Y . Zhong, H.-K. Lam, J. Han, G. Ouyang, X. Li, and H. Liu, “Appearance-based gaze estimation for asd diagnosis,”IEEE Transactions on Cybernetics, vol. 52, no. 7, pp. 6504–6517, 2022

  2. [2]

    Detection of malpractice in e-exams by head pose and gaze estimation,

    C. S. Indi, V . Pritham, V . Acharya, and K. Prakasha, “Detection of malpractice in e-exams by head pose and gaze estimation,”International Journal of Emerg- ing Technologies in Learning (Online), vol. 16, no. 8, p. 47, 2021

  3. [3]

    Experimental road safety study of the actual driver reaction to the street ads using eye tracking, multiple linear regression and decision trees methods,

    S. AlKheder, “Experimental road safety study of the actual driver reaction to the street ads using eye tracking, multiple linear regression and decision trees methods,” Expert Systems with Applications, vol. 252, p. 124222, 2024

  4. [4]

    Deep face profiler (defap): Towards explicit, non-restrained, non-invasive, facial and gaze comprehension,

    W. Khan, L. Topham, H. Alsmadi, A. Al Kafri, and H. Kolivand, “Deep face profiler (defap): Towards explicit, non-restrained, non-invasive, facial and gaze comprehension,” Expert Systems with Applications , vol. 254, p. 124425, 2024

  5. [5]

    Local eye-net: An attention based deep learning architecture for localization of eyes,

    S. Maiti and A. Gupta, “Local eye-net: An attention based deep learning architecture for localization of eyes,” Expert Systems with Applications , vol. 239, p. 122416, 2024

  6. [6]

    Eye-model-based gaze estimation by rgb-d camera,

    L. Jianfeng and L. Shigang, “Eye-model-based gaze estimation by rgb-d camera,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014, pp. 592–596

  7. [7]

    3d eye model-based gaze estimation from a depth sensor,

    X. Zhou, H. Cai, Z. Shao, H. Yu, and H. Liu, “3d eye model-based gaze estimation from a depth sensor,” in 2016 IEEE international conference on robotics and biomimetics (ROBIO). IEEE, 2016, pp. 369–374

  8. [8]

    Two-eye model- based gaze estimation from a kinect sensor,

    X. Zhou, H. Cai, Y . Li, and H. Liu, “Two-eye model- based gaze estimation from a kinect sensor,” in 2017 IEEE International Conference on Robotics and Au- tomation (ICRA). IEEE, 2017, pp. 1646–1653

Show all 43 references
  1. [9]

    Eye-gaze estimation under var- ious head positions and iris states,

    R. Jafari and D. Ziou, “Eye-gaze estimation under var- ious head positions and iris states,” Expert Systems with Applications, vol. 42, no. 1, pp. 510–518, 2015

  2. [10]

    Appearance-based gaze estimation in the wild,

    X. Zhang, Y . Sugano, M. Fritz, and A. Bulling, “Appearance-based gaze estimation in the wild,” in Proceedings of the IEEE conference on computer vi- sion and pattern recognition, 2015, pp. 4511–4520

  3. [11]

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

    ——, “Mpiigaze: Real-world dataset and deep appearance-based gaze estimation,” IEEE transac- tions on pattern analysis and machine intelligence , vol. 41, no. 1, pp. 162–175, 2017

  4. [12]

    Appearance-based gaze estimation via evaluation-guided asymmetric re- gression,

    Y . Cheng, F. Lu, and X. Zhang, “Appearance-based gaze estimation via evaluation-guided asymmetric re- gression,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 100–115

  5. [13]

    Puregaze: Purify- ing gaze feature for generalizable gaze estimation,

    Y . Cheng, Y . Bao, and F. Lu, “Puregaze: Purify- ing gaze feature for generalizable gaze estimation,” in Proceedings of the AAAI Conference on Artificial In- telligence, vol. 36, no. 1, 2022, pp. 436–443

  6. [14]

    Efe: End-to-end frame-to-gaze estima- tion,

    H. Balim, S. Park, X. Wang, X. Zhang, and O. Hilliges, “Efe: End-to-end frame-to-gaze estima- tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 2687–2696

  7. [15]

    Rotation- constrained cross-view feature fusion for multi-view appearance-based gaze estimation,

    Y . Hisadome, T. Wu, J. Qin, and Y . Sugano, “Rotation- constrained cross-view feature fusion for multi-view appearance-based gaze estimation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 5985–5994

  8. [16]

    Eg-net: Appearance-based eye gaze estimation using an efficient gaze network with attention mechanism,

    X. Wu, L. Li, H. Zhu, G. Zhou, L. Li, F. Su, S. He, Y . Wang, and X. Long, “Eg-net: Appearance-based eye gaze estimation using an efficient gaze network with attention mechanism,” Expert Systems with Ap- plications, p. 122363, 2023

  9. [17]

    Searching for mobilenetv3,

    A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y . Zhu, R. Pang, V . Vasudevan et al., “Searching for mobilenetv3,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314–1324

  10. [18]

    Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features,

    S. N. Wadekar and A. Chaurasia, “Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features,” arXiv preprint arXiv:2209.15159, 2022

  11. [19]

    Fr- net: A light-weight fft residual net for gaze estima- tion,

    T. Xu, B. Wu, R. Fan, Y . Zhou, and D. Huang, “Fr- net: A light-weight fft residual net for gaze estima- tion,” arXiv preprint arXiv:2305.11875, 2023

  12. [20]

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

    X. Zhang, Y . Sugano, M. Fritz, and A. Bulling, “It’s written all over your face: Full-face appearance-based gaze estimation,” in Proceedings of the IEEE con- ference on computer vision and pattern recognition workshops, 2017, pp. 51–60

  13. [21]

    L2cs-net: Fine-grained gaze estimation in unconstrained environments,

    A. A. Abdelrahman, T. Hempel, A. Khalifa, A. Al- Hamadi, and L. Dinges, “L2cs-net: Fine-grained gaze estimation in unconstrained environments,” in 2023 8th International Conference on Frontiers of Signal Processing (ICFSP). IEEE, 2023, pp. 98–102

  14. [22]

    Source- free adaptive gaze estimation by uncertainty reduc- tion,

    X. Cai, J. Zeng, S. Shan, and X. Chen, “Source- free adaptive gaze estimation by uncertainty reduc- tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 22 035–22 045

  15. [23]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural infor- mation processing systems, vol. 30, 2017

  16. [24]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 7132–7141

  17. [25]

    Gated channel transformation for visual recognition,

    Z. Yang, L. Zhu, Y . Wu, and Y . Yang, “Gated channel transformation for visual recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 794–11 803

  18. [26]

    Matrix cap- sules with em routing,

    G. E. Hinton, S. Sabour, and N. Frosst, “Matrix cap- sules with em routing,” inInternational conference on learning representations, 2018

  19. [27]

    Mirn: A multi-interest retrieval network with sequence-to-interest em routing,

    X. Zhang, J. Liu, S. Chang, P. Gong, Z. Wu, and B. Han, “Mirn: A multi-interest retrieval network with sequence-to-interest em routing,” Plos one , vol. 18, no. 2, p. e0281275, 2023

  20. [28]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

  21. [29]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

  22. [30]

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

    X. Zhang, S. Park, T. Beeler, D. Bradley, S. Tang, and O. Hilliges, “Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze varia- tion,” in Computer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part...

  23. [31]

    Gaze360: Physically unconstrained gaze estimation in the wild,

    P. Kellnhofer, A. Recasens, S. Stent, W. Matusik, and A. Torralba, “Gaze360: Physically unconstrained gaze estimation in the wild,” in Proceedings of the IEEE/CVF international conference on computer vi- sion, 2019, pp. 6912–6921

  24. [32]

    Rt-gene: Real-time eye gaze estimation in natural environ- ments,

    T. Fischer, H. J. Chang, and Y . Demiris, “Rt-gene: Real-time eye gaze estimation in natural environ- ments,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 334–352

  25. [33]

    Searching efficient neural architecture with multi-resolution fusion transformer for appearance-based gaze estimation,

    V . Nagpure and K. Okuma, “Searching efficient neural architecture with multi-resolution fusion transformer for appearance-based gaze estimation,” in Proceed- ings of the IEEE/CVF winter conference on applica- tions of computer vision, 2023, pp. 890–899

  26. [34]

    Gaze-swin: Enhancing gaze estimation with a hybrid cnn-transformer network and dropkey mechanism,

    R. Zhao, Y . Wang, S. Luo, S. Shou, and P. Tang, “Gaze-swin: Enhancing gaze estimation with a hybrid cnn-transformer network and dropkey mechanism,” Electronics, vol. 13, no. 2, p. 328, 2024

  27. [35]

    Gaze estimation using trans- former,

    Y . Cheng and F. Lu, “Gaze estimation using trans- former,” in 2022 26th International Conference on Pattern Recognition (ICPR). IEEE, 2022, pp. 3341– 3347

  28. [36]

    Appearance-based gaze esti- mation using dilated-convolutions,

    Z. Chen and B. E. Shi, “Appearance-based gaze esti- mation using dilated-convolutions,” in Asian Confer- ence on Computer Vision. Springer, 2018, pp. 309– 324

  29. [37]

    A coarse-to-fine adaptive network for appearance-based gaze estimation,

    Y . Cheng, S. Huang, F. Wang, C. Qian, and F. Lu, “A coarse-to-fine adaptive network for appearance-based gaze estimation,” in Proceedings of the AAAI Confer- ence on Artificial Intelligence , vol. 34, no. 07, 2020, pp. 10 623–10 630

  30. [38]

    Towards self-supervised gaze estimation,

    A. Farkhondeh, C. Palmero, S. Scardapane, and S. Es- calera, “Towards self-supervised gaze estimation,” arXiv preprint arXiv:2203.10974, 2022

  31. [39]

    Self-attention with convolution and deconvolution for efficient eye gaze estimation from a full face image,

    J. O Oh, H. J. Chang, and S.-I. Choi, “Self-attention with convolution and deconvolution for efficient eye gaze estimation from a full face image,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4992–5000

  32. [40]

    Gazecaps: Gaze estimation with self-attention-routed capsules,

    H. Wang, J. O. Oh, H. J. Chang, J. H. Na, M. Tae, Z. Zhang, and S.-I. Choi, “Gazecaps: Gaze estimation with self-attention-routed capsules,” inProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, 2023, pp. 2668–2676

  33. [41]

    Gaze es- timation via strip pooling and multi-criss-cross atten- tion networks,

    C. Yan, W. Pan, C. Xu, S. Dai, and X. Li, “Gaze es- timation via strip pooling and multi-criss-cross atten- tion networks,” Applied Sciences, vol. 13, no. 10, p. 5901, 2023

  34. [42]

    Scaling up your kernels to 31x31: Revisiting large kernel design in cnns,

    X. Ding, X. Zhang, J. Han, and G. Ding, “Scaling up your kernels to 31x31: Revisiting large kernel design in cnns,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 963–11 975

  35. [43]

    Cbam: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” in Proceed- ings of the European conference on computer vision (ECCV), 2018, pp. 3–19

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.