acceptodds
Withdrawn from ICLR 2027

GazeMixup: Direction-Magnitude Aware Feature Interpolation for Cross-Domain Gaze Estimation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Appearance-based gaze estimation often suffers from poor cross-domain generalization due to variations in identity, illumination, imaging devices, and other domain-specific factors. We introduce GazeMixup, a geometry-aware feature-level augmentation method for domain-generalized gaze estimation. The key idea is to decompose gaze features into direction and magnitude components, motivated by the observed association between feature magnitude and prediction error. GazeMixup separately interpolates these components with magnitude-aware coefficients and applies the same interpolation principle to gaze labels, enabling adaptive feature synthesis while maintaining feature–label consistency. The method is parameter-free, requires no target-domain data, and can be seamlessly integrated into existing gaze estimation networks. Extensive experiments across multiple cross-domain tasks and backbone architectures show that GazeMixup consistently outperforms vanilla Mixup and improves the generalization of baseline gaze estimators on unseen domains, demonstrating the effectiveness of geometry-aware feature interpolation for cross-domain gaze estimation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Gaze Estimetion, Domain Generalized, Cross-Domain

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bZDpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GazeMixup, a feature-level augmentation method for cross-domain gaze estimation. Motivated by the observed relationship between feature magnitude and prediction error, the method separately interpolates feature direction and magnitude, and applies a similar strategy to gaze labels. GazeMixup introduces no additional trainable parameters or target-domain data. Experiments across four backbone architectures and four cross-domain settings report improvements over vanilla Mixup, with additional ablations examining feature geometry and label interpolation.

Critical strengths

1. Simple and practical design.

GazeMixup is a lightweight, parameter-free augmentation strategy that can be integrated into existing gaze estimation architectures without additional networks or target-domain data.

2. Consistent improvements over vanilla Mixup.

The method reports improvements over vanilla Mixup across four backbones and four cross-domain settings. The ablation studies also examine the contributions of feature and label interpolation.

Critical weaknesses

1. Potential inconsistency in the Error Stability (ES) metric.

Eq. 13 defines ES as the mean absolute difference of per-sample errors. However, all 24 ES values in Table 2 exactly match the absolute differences between aggregate MAEs. These are mathematically different quantities. This raises serious concerns about the correctness of the reported ES results and the corresponding claims of cross-domain stability.

2. The claimed feature–label consistency is not sufficiently justified.

GazeMixup computes feature and label interpolation coefficients independently from their respective norms, without establishing a relationship between them. Even for a linear regression head that perfectly fits the original samples, this can introduce additional residuals on mixed samples. The improvements may therefore arise from structured target perturbations rather than the proposed geometric consistency. Table 4 highlights the importance of this distinction, yet the paper lacks controlled experiments to isolate these effects.

3. The validity of arbitrary interpolation for gaze regression remains unclear.

Randomly mixing samples with substantially different gaze directions does not guarantee that the resulting feature corresponds to the interpolated label. Prior work such as C-Mixup (NeurIPS 2022) addresses this issue through label-aware sampling, but no such comparison is provided. Furthermore, the physical meaning of gaze-label magnitude is unclear, particularly if labels are represented as yaw–pitch angles rather than 3D unit vectors.

4. Weak baselines and insufficient statistical validation.

The reported GazeTR baseline (11.93°/12.87° on Gaze360 to MPIIFaceGaze/EyeDiap) is substantially worse than previously published results (7.96°/8.88°). Although evaluation protocols may differ, this discrepancy requires explanation. Moreover, several improvements over vanilla Mixup are only 0.1–0.4°, without results across independent training seeds. These issues weaken the evidence for consistent generalization improvements.

Questions for authors

1. Can the authors verify the ES calculation?

Why do all 24 ES values match the absolute differences between aggregate MAEs rather than the per-sample metric defined in Eq. 13?

2. Can the authors isolate the actual source of GazeMixup's improvements?

Please consider two controlled baselines: (a) standard linear feature–label Mixup at the same feature layer, and (b) matched target perturbation without the proposed feature interpolation. These would help distinguish geometric regularization from supervision perturbation.

3. What is the exact gaze-label representation?

Are labels yaw–pitch angles or 3D gaze vectors? What is the physical meaning of their magnitude, and how are zero norms handled? Have the authors considered spherical interpolation or gaze-angle-aware sample pairing?

4. Can the authors clarify the baseline discrepancies and statistical reliability?

Why is the reported GazeTR baseline substantially worse than published results? Please clarify the evaluation protocol and report performance across multiple independent training seeds.

Reviewer_qJ4sOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GazeMixup, a parameter-free feature augmentation method designed to improve cross-domain gaze estimation. Based on the observation that feature magnitude is strongly correlated with prediction error, the method decomposes gaze features into magnitude and direction, interpolates them separately using magnitude-aware coefficients, and applies this identical interpolation scheme to the corresponding gaze labels.

Empirical validation shows that GazeMixup consistently improves various baselines (VGG, ResNet18/50, GazeTR) across four major benchmarks in both cross-domain and in-domain settings.

However, the theoretical gaps, methodological ambiguities, weak visual evidence, and presentation issues must be rigorously addressed.

Critical strengths
  1. A feature augmentation method is proposed. The method is plug-and-play and parameter-free.

  2. Experimental validation is provided. Under the cross-dataset and inter-dataset experimental settings on four datasets, the proposed GazeMixUp method can improve the gaze estimation performance of the following baseline methods: VGG, ResNet18, ResNet50, and GazeTR. Ablation experiments on different feature interpolation methods and different label mixing methods are provided.

Critical weaknesses

1. Lack of Theoretical Support for the Feature-Label Consistency Assumption: One of the core operations of the proposed method is decoupling feature magnitude and direction for separate interpolation, followed by applying the exact same interpolation scheme to the gaze labels (Eq. 9-11). However, enforcing this geometric consistency between features and labels appears to be a strong, heuristic assumption. Although the ablation study (Table 4) demonstrates empirical benefits, it reads as though the regressor head is merely being forced to overfit this artificially constructed (feature, label) mapping. The paper lacks a physical or theoretical explanation for why applying a magnitude-aware non-linear scaling to the labels translates to physically meaningful gaze directions in the real world.

2. Logical Gap Between Motivation and Method (Explicit Constraint vs. Data Augmentation): In Section 4.1, the authors conclude that prediction errors are highly correlated with feature magnitude, and that the magnitude inherently contains error-related information. However, there is a logical leap from this observation to the final method design. If feature magnitude is the primary source of error, why not introduce a regularization loss penalty during training to explicitly constrain or align the magnitudes to a fixed range, rather than relying on magnitude-based augmentation? The authors should clarify the fundamental advantages of magnitude-aware interpolation over explicit magnitude regularization.

3. Weak Empirical Visual Support for the Direction-Gaze Correlation Claim: The authors repeatedly claim that "feature direction has a stronger relationship with gaze variation" as a core motivation for their method design. However, the visualizations provided in Figure 2 fail to convincingly support this assertion. Specifically, in the ResNet18 baseline plots (top row, the first two columns labeled "Angle"), the colors representing the 1D directional projection are chaotically mixed and scattered across the yaw and pitch axes. There is no significant, discernible pattern or clear gradient that visually demonstrates a strong alignment between the feature direction and gaze variation. This lack of clear visual evidence weakens the empirical foundation of the proposed feature decomposition strategy.

4. Ambiguity in Gaze Direction Representation and Its Impact: The paper does not explicitly specify whether the base model's output (and the corresponding label yy) is represented as Euler angles (e.g., 2D pitch and yaw) or as 3D Cartesian directional vectors (x,y,z)(x, y, z). Different mathematical representations of the gaze direction possess distinct geometric properties, which would naturally alter the magnitude and angular distributions of the corresponding model features. Consequently, this representational choice might significantly impact the effectiveness and mathematical validity of GazeMixup (especially regarding the magnitude-aware label interpolation mechanism in Eq. 9-11). The authors should clarify the exact representation used and discuss how different coordinate systems influence the behavior and performance of the proposed method.

5. Limitations in Out-of-Distribution (OOD) Generalization: As a source-only domain generalization method, GazeMixup essentially relies on feature augmentation to increase sample diversity. However, this interpolation is inherently bounded by the "convex hull" of the feature magnitudes and angles present in the source domain. For the method to perform well on an unknown test domain, it assumes that the test features will fall within the manifold constructed by the source features. How significant are the differences in feature magnitude and angle across different domains for the same gaze estimator? Does the proposed method possess any extrapolation capability for OOD samples that fall outside this augmented range?

6. Poor Presentation and Figure Clarity:

Several core figures lack adequate explanations, severely hindering readability:

  • Figure 1 (Motivation) is difficult to follow: It contains numerous subfigures and elements without sufficient in-text explanation of their logical relationships. For instance, what is the exact physical meaning of the colorful distribution maps on the left side of Fig. 1(a)? What do "Regular Distribution" and "Feature Regularization" specifically refer to? How should a reader visually interpret "substantial cross-domain error" and its "reduced" state from these plots?

  • Figure 2 lacks clear descriptions: The caption and main text fail to define the axes and data points clearly. What do the individual points represent within the (Yaw, Pitch) coordinates? What do "Angle" (column 1) and "Length" (column 2) specifically refer to (e.g., does "Angle" refer to ψ\psi as defined in Eq. 3)? Furthermore, the FSCI method is abruptly introduced here without adequate contextual setup.

  • Figure 6 is heavily illegible: The scatter points, legends, and axis text are far too small. Even when zoomed in digitally, it is difficult to distinguish the clusters, which weakens the qualitative claims made by the authors.

Questions for authors
  1. Clarification on Output Representation: Please explicitly specify the format of the label yy used in your experiments (i.e., 2D angles or 3D vectors). As noted in Weakness 3, this is crucial for assessing the validity of the label interpolation formulas.

  2. In-dataset Evaluation for GazeTR: In Table 6, you report the in-dataset results for ResNet and VGG but omit GazeTR. Given that GazeTR demonstrates competitive performance in the cross-dataset evaluations, could you provide its in-dataset performance on ETH-XGaze and Gaze360? Additionally, does GazeMixup operate differently on CNN-based architectures versus Transformer-based ones?

  3. Restructuring the Theory Section: The Theory section in Appendix A is highly dense with derivations. I recommend stating the core theoretical conclusions (theorems/corollaries) first, followed by the detailed mathematical proofs. This structure will help readers grasp the main theoretical takeaways much more efficiently.

  4. Figure Revisions: Please provide revised versions or detailed captions for Figures 1 and 2 during the rebuttal phase, and significantly improve the legibility and font size of Figure 6.

Reviewer_Ba55OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript addresses cross-domain appearance-based gaze estimation and proposes a feature-level augmentation that interpolates the direction and magnitude of penultimate features separately, with a magnitude-aware coefficient. Experiments cover four transfer settings built from ETH-XGaze, Gaze360, MPIIFaceGaze, and EyeDiap, four backbones (VGG, ResNet18, ResNet50, GazeTR), a comparison with cross-domain gaze methods, and ablations of the feature and label interpolation variants.

Critical strengths
  1. The method is simple and parameter-free, and the algebra is internally consistent: the product form in Eq. 6 and the linear form in Eqs. 7 to 8 coincide, and the coefficient behaviour described in the text follows from them.
  2. The evaluation crosses four backbones with four transfer settings, and all sixteen cells improve over vanilla Mixup, which is more than the usual single-backbone demonstration.
  3. Ablations decompose the design, comparing magnitude-only, direction-only, and normalized feature interpolation in Table 3 and three label constructions in Table 4.
  4. The manuscript reports a source-domain sensitivity metric and an angular error profile in Figure 7, both testing whether the gain is concentrated in a narrow regime.
Critical weaknesses
  1. The label interpolation is not well defined, and the ablation that is supposed to support it cannot be explained by the stated principle. The method writes the label as a vector y and interpolates its magnitude and direction through Eqs. 9 to 11, but the manuscript never states what y is. If y is a gaze direction it is a unit vector, so both label magnitudes equal one, the scheme collapses to convex mixing, and the two rows of Table 4 must coincide; instead they differ by 3.0 to 3.7 degrees. If y is a yaw-pitch pair in radians, then the norm of y is not a geometric magnitude of gaze, and Eqs. 9 to 11 reweight samples by the inverse of that norm without justification, so the operation no longer corresponds to the direction-magnitude decomposition that motivates the feature side. Either reading leaves the central claim of the third contribution, that a shared magnitude-aware principle in feature and label space is what helps, unsupported by its own evidence. The tensor shape of y should be stated, the two constructions derived, and a numerical example given.
  2. Every result comes from a single training run with no seed variance or interval, and several of the claimed gains are smaller than plausible run-to-run variation. On the DE to DD setting, ResNet18 improves from 7.83 to 7.67 degrees over the unaugmented baseline, and ResNet50 from 8.90 to 7.45; the former is 0.16 degrees. Three seeds with paired intervals per setting are needed before the smaller margins can be read as effects.
  3. The theory in Appendix A does not distinguish the proposed interpolation from the alternatives it is meant to beat. The bound in Eq. 23 follows the standard mixup analysis template and applies to any convex interpolation of features under the stated Lipschitz and quadratic assumptions; it is never instantiated for the magnitude-aware path, nor compared with the bound for plain Mixup, so it cannot explain the reported differences. The appendix also assumes a uniform mixing coefficient while Section 5.1 states a Beta distribution, and it analyzes only the feature side, leaving the label construction that Table 4 leans on without any analysis.
Questions for authors
  1. Are the competing methods in Table 2 retrained under the 80-epoch protocol of Section 6.2, and what are their hyperparameters?
  2. What is the correlation between Error Stability and target-domain error across the methods in Table 2, and does it support using ES as evidence of robustness?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.