Research map / Label-efficient and robust learning / Domain adaptation and generalization
Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Gaze Label Alignment (GLA), a pre-processing procedure that aligns per-domain gaze labels to an anchor domain via a fitted transformation so that gaze estimators trained on aligned labels generalize better across datasets.
Problem
Cross-domain gaze estimation degrades because, besides data distribution shift, gaze labels across datasets (and individuals) are biased by calibration of acquisition systems and physiological differences, which existing domain generalization/adaptation methods ignore.
Approach
Train a feature extractor on all domains, fit a regressor on one anchor domain, and learn a rigid-transform-based mapping that aligns each other domain's ground-truth labels to the anchor's predicted labels; retrain on aligned labels. The method is model-agnostic and can be plugged into any DG/DA gaze method.
Key results
- GLA on ETH-XGaze+GazeCapture training reduces average cross-domain angular error from 7.11° to 6.76° for the baseline (MPIIFaceGaze 6.83→6.58, EyeDiap 7.39→6.93).
- Plugging GLA into SOTA methods lowers errors by about 3-8% relative, e.g. UnReGA average 5.56°→5.38° and fine-tuning average 5.43°→5.01°.
- Using three source domains (ETH-XGaze, Gaze360, GazeCapture) with GLA gives average error 5.30° versus 5.60° without GLA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply GLA label alignment to the baseline model | Baseline trained on ETH-XGaze + GazeCapture · MPIIFaceGaze | angular error (°): 6.83 -> 6.58 | Helped |
| Apply GLA label alignment to the baseline model | Baseline trained on ETH-XGaze + GazeCapture · EyeDiap | angular error (°): 7.39 -> 6.93 | Helped |
| Train CLIP-Gaze on GLA-aligned labels | CLIP-Gaze · MPIIFaceGaze | angular error (°): 6.36 -> 6.13 | Helped |
| Train CLIP-Gaze on GLA-aligned labels | CLIP-Gaze · EyeDiap | angular error (°): 6.45 -> 6.30 | Helped |
| Train UnReGA on GLA-aligned labels | UnReGA · MPIIFaceGaze | angular error (°): 5.27 -> 5.05 | Helped |
| Train UnReGA on GLA-aligned labels | UnReGA · EyeDiap | angular error (°): 5.86 -> 5.71 | Helped |
| Use the gaze-target rotation and translation alignment function rather than offset, linear, affine, or homography mappings (ablation) | GLA · MPIIFaceGaze + EyeDiap average | angular error (°): 6.95, 6.87, 6.81, 6.80 -> 6.76 | Helped |
| Align each person's data as a separate unit rather than aligning each dataset as a unit (ablation) | GLA · MPIIFaceGaze + EyeDiap average, ETH-XGaze + GazeCapture sources | angular error (°): 6.96 -> 6.76 | Helped |
| Align each dataset as a unit rather than using no label alignment (ablation) | GLA · MPIIFaceGaze + EyeDiap average, ETH-XGaze + GazeCapture sources | angular error (°): 7.06 -> 6.96 | Helped |
| Select ETH-XGaze, Gaze360, or GazeCapture as the anchor domain (ablation) | GLA with ETH-XGaze + Gaze360 + GazeCapture sources · MPIIFaceGaze + EyeDiap average | angular error (°): 5.30, 5.33, 5.48 | No clear effect |
| Increase the number of complete alignment passes from one to two (ablation) | GLA · MPIIFaceGaze + EyeDiap average | angular error (°): 6.76 -> 6.72 | No clear effect |
| Increase the number of complete alignment passes from one to three (ablation) | GLA · MPIIFaceGaze + EyeDiap average | angular error (°): 6.76 -> 6.74 | No clear effect |
Surprising findings
- Adding a second training dataset (ETH-XGaze + GazeCapture) without alignment can yield higher error on MPIIFaceGaze than training on GazeCapture alone (6.83° vs 6.72°), which the authors attribute to label shift.
Limitations
- Gains on the in-domain alignment rely on an anchor domain choice; the paper notes some variation with the selected domain.
- Alignment assumes calibration error mainly comes from the gaze target and uses a shared rigid transform, a simplified model of label deviation.
- Evaluation is limited to two target datasets (MPIIFaceGaze, EyeDiap) and a ResNet-18 backbone.
Experimental setup
Datasets: ETH-XGaze Gaze360 GazeCapture MPIIFaceGaze EyeDiap
Benchmarks: MPIIFaceGaze EyeDiap
Metrics: angular error (degrees)
Baselines: Baseline (ResNet-18 gaze model) PureGaze CDG CLIP-Gaze PnP-GA RUDA CSA UnReGA
Builds on: PureGaze CLIP-Gaze PnP-GA
Closest papers by meaning
- A Generalized Label Shift Perspective for Cross-Domain Gaze Estimation (NeurIPS 2025 · similarity 0.91)
- Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process (AAAI 2024 · similarity 0.89)
- Suppressing Uncertainty in Gaze Estimation (AAAI 2024 · similarity 0.89)
- From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation (CVPR 2024 · similarity 0.87)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.86)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.86)
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.