Research map / Label-efficient and robust learning / Recognition and long-tail
Suppressing Uncertainty in 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 SUGE, which estimates image and label uncertainty via a triplet-label consistency (pseudo, ground-truth, and neighboring labels) and uses GMM-based confidences for sample weighting and label correction in gaze estimation training.
Problem
Gaze estimation training is hurt by low-quality images (occlusion, blur, non-face images) and mislabeled gaze targets from imperfect fixation-based annotation.
Approach
Builds a neighboring label from feature-space K-NN with locally linear reconstruction weights, then measures discrepancies among ground-truth, pseudo, and neighboring labels (TupleMD for label quality, TripleMD for image quality); a bimodal GMM turns these into confidences used to weight samples and correct labels, trained with two co-training networks.
Key results
- SUGE (GazeTR) reaches 4.01/10.51/5.04 degrees angle error on MPII/Gaze360/EyeDiap, improving GazeTR† by 0.10 and 0.30 degrees on Gaze360 and EyeDiap.
- SUGE (Gaze360) reaches 4.07/10.52/5.05 degrees, improving the re-implemented Gaze360† by 0.26 and 0.41 degrees on Gaze360 and EyeDiap.
- On EyeDiap, SUGE (5.05) outperforms Co-Teaching (5.18) and DivideMix (5.16) applied to the Gaze360 baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SUGE to Gaze360 | Gaze360 · Gaze360 | angle error (°): 10.78 -> 10.52 | Helped |
| Apply SUGE to Gaze360 | Gaze360 · EyeDiap | angle error (°): 5.46 -> 5.05 | Helped |
| Apply SUGE to GazeTR | GazeTR · MPIIFaceGaze | angle error (°): 4.00 -> 4.01 | No clear effect |
| Apply SUGE to GazeTR | GazeTR · Gaze360 | angle error (°): 10.61 -> 10.51 | Helped |
| Apply SUGE to GazeTR | GazeTR · EyeDiap | angle error (°): 5.34 -> 5.04 | Helped |
| Add neighboring labeling (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.16 -> 5.05 | Helped |
| Add reconstruction weighting for neighboring labels (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.10 -> 5.05 | Helped |
| Add sample weighting (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.10 -> 5.05 | Helped |
| Add label correction (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.17 -> 5.05 | Helped |
| Add sample weighting and label correction (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.23 -> 5.05 | Helped |
| Add co-training instead of separate self-training (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.14 -> 5.05 | Helped |
| Use the entire label correction composition (ablation) | SUGE with Gaze360 · EyeDiap | angle error (°): 5.10 -> 5.05 | Helped |
Surprising findings
- On MPIIFaceGaze, SUGE gives no improvement for the GazeTR backbone (4.00 to 4.01) and is slightly worse for the Gaze360 backbone (4.17 to 4.07 is actually better, but the GazeTR case is flat or slightly worse).
Limitations
- Gains are small (about 0.1 to 0.4 degrees) and no gain on MPIIFaceGaze with the GazeTR backbone.
- Relies on hand-set thresholds (tau = 0.5), epsilon, and K, with K and thresholds tuned on EyeDiap.
- Compared with noisy-label methods only on the Gaze360 baseline and EyeDiap, and not on MPIIFaceGaze.
Experimental setup
Datasets: EyeDiap MPIIFaceGaze Gaze360 ETH-XGaze
Benchmarks: MPIIFaceGaze Gaze360 EyeDiap
Metrics: angle error (degrees)
Baselines: Full Face CA-Net Gaze360 CADSE GazeTR Co-Teaching DivideMix
Builds on: DivideMix Co-Teaching Gaze360 GazeTR
Closest papers by meaning
- Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation (AAAI 2025 · similarity 0.89)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.87)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.87)
- A Generalized Label Shift Perspective for Cross-Domain Gaze Estimation (NeurIPS 2025 · similarity 0.86)
- UVAGaze: Unsupervised 1-to-2 Views Adaptation for Gaze Estimation (AAAI 2024 · similarity 0.86)
- De^2Gaze: Deformable and Decoupled Representation Learning for 3D Gaze Estimation (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.