Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Generalized Label Shift Perspective for Cross-Domain 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.
Reframes cross-domain gaze estimation as a generalized label shift problem and proposes GLSGE, which reweights source labels with a truncated Gaussian and aligns conditionals via a probability-aware conditional operator discrepancy (PCOD).
Problem
Cross-domain gaze estimation suffers from domain shift; existing domain-invariant feature methods are argued to be insufficient under label shift, and existing label-shift correction methods assume discrete classes, which does not fit continuous gaze labels.
Approach
Model the gaze label distribution as a bivariate truncated Gaussian to compute continuous importance weights for label shift correction, and derive a probability-aware conditional operator discrepancy (PCOD) so the reweighted source label distribution can be embedded in conditional invariant learning, jointly forming a GLS correction framework.
Key results
- With ResNet-18 backbone, GLSGE achieves 6.06 average angular error versus 8.33 for the ResNet-18 baseline (27.2% reduction) and beats prior methods on average (best prior PnP-GA+ at 6.20).
- Plugged into GazeTR, GLSGE reduces average error by 19.1% (8.99 to 7.27); plugged into FSCI, it reduces average error by 12.1% (6.95 to 6.11).
- With ResNet-50, GLSGE reaches 6.01 average error, a 26.3% reduction from 8.16, with the best average among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply GLSGE to ResNet-18 | ResNet-18 · Average across four CDGE tasks | angular error (degrees): 8.33 -> 6.06; 27.2% | Helped |
| Apply GLSGE to ResNet-50 | ResNet-50 · Average across four CDGE tasks | angular error (degrees): 8.16 -> 6.01; 26.3% | Helped |
| Apply GLSGE to GazeTR | GazeTR · Average across four CDGE tasks | angular error (degrees): 8.99 -> 7.27; 19.1% | Helped |
| Apply GLSGE to FSCI | FSCI · Average across four CDGE tasks | angular error (degrees): 6.95 -> 6.11; 12.1% | Helped |
| Add label shift correction to source loss (ablation) | ResNet-18 CDGE · ETH-XGaze → EyeDiap | angular error (degrees): 9.03 -> 6.70 | Helped |
| Add conditional alignment with COD (ablation) | ResNet-18 CDGE · ETH-XGaze → EyeDiap | angular error (degrees): 9.03 -> 7.42 | Helped |
| Replace COD with probability-aware PCOD (ablation) | Source loss + conditional alignment · ETH-XGaze → EyeDiap | angular error (degrees): 7.42 -> 7.15 | Helped |
| Combine label shift correction with COD conditional alignment (ablation) | ResNet-18 CDGE · ETH-XGaze → EyeDiap | angular error (degrees): 9.03 -> 6.32 | Helped |
| Replace COD with PCOD in the label-reweighted method (ablation) | Weighted source loss + conditional alignment · ETH-XGaze → EyeDiap | angular error (degrees): 6.32 -> 6.21 | Helped |
| Use only 100 unlabeled target samples (ablation) | GLSGE with ResNet-18 · Average across four CDGE tasks | angular error (degrees): 6.06 -> 6.23 | Hurt |
Surprising findings
- FSCI+GLSGE has larger error than ResNet-18+GLSGE on some settings, which the authors attribute to loss of task-specific factors during DG training.
- Using only 100 unlabeled target samples still gives 6.23 average error, close to the 6.06 obtained with over 1000 samples.
Limitations
- Requires target-domain samples during training, limiting use on unseen domains.
- Kernel-based conditional alignment is computationally intensive.
- Truncated Gaussian label model is a simple prior and may not fit multimodal gaze distributions such as driver gaze.
Experimental setup
Datasets: ETH-XGaze Gaze360 MPIIFaceGaze EyeDiap
Benchmarks: ETH-XGaze to MPIIFaceGaze ETH-XGaze to EyeDiap Gaze360 to MPIIFaceGaze Gaze360 to EyeDiap
Metrics: angular error (degrees)
Baselines: PureGaze GazeCon GFAL AGG FSCI CGaG DAGEN PnP-GA GSA-Gaze HFC DCUA PnP-GA+
Builds on: Conditional Operator Discrepancy (COD) DAGE-GRAM PnP-GA+ GazeTR FSCI
Closest papers by meaning
- Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation (AAAI 2025 · similarity 0.91)
- Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process (AAAI 2024 · similarity 0.87)
- From Feature to Gaze: A Generalizable Replacement of Linear Layer for Gaze Estimation (CVPR 2024 · similarity 0.87)
- Suppressing Uncertainty in Gaze Estimation (AAAI 2024 · similarity 0.86)
- CLIP-Gaze: Towards General Gaze Estimation via Visual-Linguistic Model (AAAI 2024 · similarity 0.85)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · 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.