Research map / Label-efficient and robust learning / Domain adaptation and generalization
Escaping the Low-Frequency Bias: Adversarial Frequency Perturbation for Generalisable Gaze Estimation
Yang Xu, Feng Lu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| randomly scale source low-frequency amplitudes during training | ResNet-18 baseline · cross-domain average, ResNet-18 | angular error (deg): 8.54 -> 7.20 | Improved |
| remove low-frequency components entirely | ResNet-18 baseline · cross-domain average, ResNet-18 | angular error (deg): 8.54 -> 14.53 | Hurt |
| add adversarial low-frequency perturbation (ALFP) | ResNet-18 · cross-domain average, ResNet-18 | angular error (deg): 8.54 -> 6.41 | Improved |
| add ALFP to ResNet-50 | ResNet-50 · cross-domain average | angular error (deg): 8.08 -> 6.15 | Improved |
| generate perturbation in spatial domain instead of frequency domain (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 6.82 -> 6.41 | Improved |
| use AdaIN statistics-level mixing instead of element-wise mixing (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 8.92 -> 6.41 | Improved |
| add Lstat batch-level statistical constraint (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 7.49 -> 6.41 | Improved |
| batch-level instead of sample-level Lstat (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 6.95 -> 6.41 | Improved |
| amplitude-domain instead of pixel-domain Lstat (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 7.26 -> 6.41 | Improved |
| range constraint instead of diversity encouragement (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 7.48 -> 6.41 | Improved |
| PGNet adversarial generation instead of within-domain images (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 6.79 -> 6.41 | Improved |
| PGNet adversarial generation instead of 300W-LP images (ablation) | ALFP · cross-domain average, ResNet-18 | angular error (deg): 7.07 -> 6.41 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PureGaze — PureGaze: Purifying Gaze Feature for Generalizable Gaze Estimation (AAAI 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Through the Frequency Lens: Cross-Domain Generalisable Gaze Estimation with Adaptive Modulation (CVPR 2026 · similarity 0.91)
- Learning a Generalized Gaze Estimator from Gaze-Consistent Feature (AAAI 2023 · similarity 0.89)
- PureGaze: Purifying Gaze Feature for Generalizable Gaze Estimation (AAAI 2022 · similarity 0.88)
- Generalizing Gaze Estimation With Outlier-Guided Collaborative Adaptation (ICCV 2021 · similarity 0.87)
- Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process (AAAI 2024 · similarity 0.85)
- From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
