Research map / 3D vision / Depth and stereo
UVAGaze: Unsupervised 1-to-2 Views Adaptation 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.
Adapts a pre-trained single-view gaze estimator to arbitrarily placed, unlabeled dual cameras via dual-view mutual supervision and a momentum head-pose stabilization loss.
Problem
Single-view gaze estimation degrades at large head angles, while existing multi-view gaze methods need costly multi-view annotations and fixed, known camera poses at test time.
Approach
Unsupervised 1-to-2 view adaptation: the two views' gaze predictions, transformed into the head coordinate system, supervise each other with the more reliable view (lower head angle) as pseudo-label, while a momentum-averaged constraint on the inter-camera rotation element stabilizes head pose predictions; pre-training data loss is added to prevent drift.
Key results
- On the Gaze360-to-ETH cross-dataset setting, UVAGaze reduces overall Mono angular error from 18.73 to 13.05 and Dual-S error to 9.93 (about 47% reduction versus the baseline Dual-S of 16.06 for select-front).
- In-dataset ETH-to-ETH, Dual-S error drops from 3.94 to 3.27 across the nine camera pairs.
- Ablation on Gaze360-to-ETH: without the stabilization module, mutual supervision alone amplifies error (Mono 57.73); with stabilization and pre-training constraint the final model reaches 13.05.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Adapt the single-view estimator to dual views using UVAGaze | ResNet18 pre-trained on Gaze360 · Gaze360→ETH (Overall) | Mono angular error: 18.73 -> 13.05 | Helped |
| Adapt the single-view estimator to dual views using UVAGaze | ResNet18 pre-trained on Gaze360 · Gaze360→ETH (Overall) | Dual-S angular error: 16.06 -> 9.93 | Helped |
| Adapt the single-view estimator to dual views using UVAGaze | ResNet18 pre-trained on ETH-MV · ETH→ETH (Overall) | Mono angular error: 4.75 -> 4.21 | Helped |
| Add mutual supervision without stabilization or pre-training gaze constraint (ablation) | UVAGaze · Gaze360→ETH | Mono angular error: 18.73 -> 57.73 | Hurt |
| Add momentum head pose stabilization to mutual supervision (ablation) | UVAGaze · Gaze360→ETH | Mono angular error: 57.73 -> 17.43 | Helped |
| Add pre-training gaze constraint to mutual supervision (ablation) | UVAGaze · Gaze360→ETH | Mono angular error: 57.73 -> 13.33 | Helped |
| Add pre-training gaze constraint to mutual supervision with head pose stabilization (ablation) | UVAGaze · Gaze360→ETH | Mono angular error: 17.43 -> 13.05 | Helped |
Surprising findings
- Naive mutual supervision without head-pose stabilization causes severe error amplification (Mono error rises to 57.73 from 18.73 baseline).
- PnP-GA suffers error amplification in the Gaze360-to-ETH cross-dataset setting (Mono 26.36 vs baseline 18.73).
Limitations
- Dual-view adaptation depends on accurate initial head pose; inaccurate pre-trained head poses reduce adaptation benefit.
- Dual-A averaging is hurt by head pose error, so select-front is preferred when head poses are unreliable.
- Evaluation uses only ETH-XGaze multi-view data with a fixed split of 18 cameras into 9 pairs.
Experimental setup
Datasets: Gaze360 ETH-XGaze
Benchmarks: ETH-XGaze (ETH-MV) Gaze360
Metrics: angular error (Mono) select-front error (Dual-S) average error (Dual-A) head pose error (HPose)
Baselines: PnP-GA RUDA DAGEN Gaze360 GazeAdv
Builds on: ResNet
Closest papers by meaning
- Unsupervised Gaze Representation Learning from Multi-view Face Images (CVPR 2024 · similarity 0.87)
- OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild (NeurIPS 2025 · similarity 0.87)
- Enhancing 3D Gaze Estimation in the Wild using Weak Supervision with Gaze Following Labels (CVPR 2025 · similarity 0.86)
- De^2Gaze: Deformable and Decoupled Representation Learning for 3D Gaze Estimation (CVPR 2025 · similarity 0.86)
- Suppressing Uncertainty in Gaze Estimation (AAAI 2024 · similarity 0.86)
- Gaze from Origin: Learning for Generalized Gaze Estimation by Embedding the Gaze Frontalization Process (AAAI 2024 · 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.