Research map / Label-efficient and robust learning / Domain adaptation and generalization
Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process
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 GFAL, a gaze domain generalization framework that embeds a gaze frontalization process (image-warping module rotating eyes to (0,0)) as an auxiliary task with a consistency loss, training on ETH-XGaze or Gaze360 without target data.
Problem
Cross-domain gaze estimation: gaze estimators trained on one dataset degrade on unseen datasets, attributed to infinite, continuous gaze regression labels.
Approach
Use gaze frontalization (rotating the eyeball in the input face image to look at the camera, (0,0)) as an auxiliary task; the feature extractor must both estimate gaze and produce a warping field that frontalizes the eyes, with a consistency loss re-estimating gaze and head orientation on the warped image.
Key results
- GFAL (ResNet-18) reaches 5.72° on ETH-XGaze-to-MPIIGaze and 6.97° on ETH-to-EyeDiap, versus 7.48° and 7.86° for the baseline.
- Gaze360-to-MPIIGaze error is 7.18° and Gaze360-to-EyeDiap is 7.38° with ResNet-18, improving over the baseline by 6.75% and 9.11%.
- Plugging image warping into the framework outperforms DeepWarp, FAZE and ST-ED gaze-redirection variants on average (6.81° vs 7.48° to 8.00°).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add gaze frontalization as an auxiliary task | ResNet-18 gaze estimation · ETH-XGaze → MPIIGaze | angular error (degrees): 7.48 -> 5.72 | Helped |
| Add gaze frontalization as an auxiliary task | ResNet-18 gaze estimation · ETH-XGaze → EyeDiap | angular error (degrees): 7.86 -> 6.97 | Helped |
| Add gaze frontalization as an auxiliary task | ResNet-18 gaze estimation · Gaze360 → MPIIGaze | angular error (degrees): 7.70 -> 7.18 | Helped |
| Add gaze frontalization as an auxiliary task | ResNet-18 gaze estimation · Gaze360 → EyeDiap | angular error (degrees): 8.12 -> 7.38 | Helped |
| Use MS-SSIM instead of L1, L2, or SSIM for the frontalization loss (ablation) | GFAL · Cross-dataset gaze estimation | angular error (degrees): L1 7.28, 7.40, 6.76, 7.47; L2 7.12, 8.25, 7.18, 8.19; SSIM 6.78, 8.59, 6.33, 7.56; MS-SSIM 5.72, 6.97, 7.18, 7.38 | Mixed |
| Use the reference gaze and head labels in consistency loss (L*_con) instead of predicted targets (L_con) (ablation) | GFAL · Cross-dataset gaze estimation | angular error (degrees): 6.02, 8.16, 7.91, 7.47 -> 5.72, 6.97, 7.18, 7.38 | Helped |
| Frontalize from (θ, ψ) to (0,0) rather than from (0,0) to (θ, ψ) (ablation) | GFAL · Cross-dataset gaze estimation | angular error (degrees): 7.33, 7.15, 7.56, 7.87 -> 5.72, 6.97, 7.18, 7.38 | Helped |
| Add consistency loss to gaze frontalization (ablation) | GFAL · Cross-dataset gaze estimation | angular error (degrees): 7.92, 8.92, 7.41, 8.59 -> 5.72, 6.97, 7.18, 7.38 | Helped |
| Replace image warping with gaze redirection (ablation) | GFAL · Cross-dataset gaze estimation | angular error (degrees): DeepWarp 7.35, 8.93, 6.73, 6.89; FAZE 8.10, 9.37, 6.87, 7.66; ST-ED 7.22, 8.20, 7.40, 8.15; warping 5.72, 6.97, 7.18, 7. | Mixed |
Surprising findings
- Using only the frontalization loss without the consistency loss does not improve cross-domain performance (average 8.21° vs baseline 7.79°).
- ResNet-50 does not always beat ResNet-18, which the authors attribute to overfitting of deeper networks.
Limitations
- Gaze360 lacks head orientation and identity labels, so SSIM and CosFace/k-means substitutes are used for reference selection.
- Gaze redirection plug-ins are complex and harder to train, and do not outperform the warping implementation on average.
Experimental setup
Datasets: ETH-XGaze Gaze360 MPIIGaze EyeDiap
Benchmarks: MPIIGaze EyeDiap
Metrics: angular error (degrees)
Baselines: Full-Face RT-Gene Dilated-Net CA-Net GazeTR PureGaze FAZE RAT LatentGaze GazeCon CDG ADDA
Builds on: Baseline (Zhang et al. 2020) PureGaze ResNet MS-SSIM
Closest papers by meaning
- From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation (CVPR 2024 · similarity 0.91)
- Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation (AAAI 2025 · similarity 0.89)
- CLIP-Gaze: Towards General Gaze Estimation via Visual-Linguistic Model (AAAI 2024 · similarity 0.89)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.88)
- A Generalized Label Shift Perspective for Cross-Domain Gaze Estimation (NeurIPS 2025 · similarity 0.87)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.87)
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.